News n°1

The AI That Steals Your Keys (Literally)

The new frontier of cybercrime is no longer a hacker in a dark room, but an SEO algorithm convincing you to download destruction right into your terminal. And the victim? Probably a proud developer who thinks they’re a genius.

There is a certain bitter irony in the fact that in 2026, the year we promise to automate everything with artificial intelligence, the most effective way to destroy a company is still getting a human being to fall for a trick. It happened again, and this time the targets are the code “geniuses” who think they’re invincible. A gang of financially motivated criminals decided that impersonating Google Gemini CLI and Anthropic’s Claude Code is the new way to make fast cash. The technique? A mix of SEO poisoning and social engineering so disarmingly simple it makes you sick to your stomach.

Picture the scene: you’re a developer, you’re in a rush, you search for “Gemini CLI” on Google. Instead of finding the official site, you end up on geminicli.co.com. It looks official, it sounds almost credible, thanks to the .co.com suffix hackers use to look legitimate without wasting a single dime on real certificates. Or maybe, if you’re particularly unlucky, you fell into the “InstallFix” trap, where the attackers actually paid for Google Ads to guarantee themselves the top spot in the rankings. Once you land on this meticulously cloned page, you’re asked to do just one thing: copy and paste a PowerShell command into your terminal. “It’s just to install the AI,” you tell yourself. “I’m a professional, I know what I’m doing.”

And here comes the plot twist. While the terminal visibly displays the legitimate npm install command that makes you feel safe, something much more sinister is happening under the hood. The script performs a masterclass in duality: on one side, the installation you see; on the other, a hidden COM object that downloads and launches a fileless infostealer right into your computer’s memory. No files to scan, no tracks on the disk—just pure code living in your RAM while you think you’re installing software. For Claude Code fans, the version is even more creative: they use mshta.exe to execute a “polyglot” file, a monstrous hybrid that contains signed, legitimate Microsoft Bing packages, but with a malicious tail attached that security controls struggle to spot.

Once inside, the malware doesn’t waste any time. It disables Windows defenses (ETW and AMSI) like they’re plastic toys and starts combing through your system with scout-like precision. It doesn’t just want your passwords; it wants your digital life. It grabs login data from Chrome, Edge, Brave, and Firefox. It takes session tokens from Slack, Teams, Discord, Zoom, and Telegram. It steals configurations from PuTTY, WinSCP, and OpenVPN. Finally, it stuffs its pockets with SSH keys, OAuth tokens, and CI/CD credentials. The beauty of it, or perhaps the worst part, is that by stealing session cookies, it completely bypasses your multi-factor authentication (MFA). That two-factor padlock that makes you feel so secure? To them, it’s as if it doesn’t even exist. All of this is then sold on the black market to those who know how to use it, turning your professional credibility into bargaining chips for access brokers.

This campaign, identified in March 2026, is not an isolated operation. The attackers have a solid infrastructure backed by “bulletproof” hosting in the Netherlands that refuses to take down their servers, and a network of over 30 domains impersonating everything from Node.js to Monero. They hit government, education, electronics, and even the food sector. The message is clear: it doesn’t matter how smart your code is, if you’re stupid enough to copy-paste commands from Google, you’ve already lost. The only real defense is to stop blindly trusting what you see at the top of the search page and start doubting anyone who asks you to copy-paste without explanation.

Don’t be like the others. Before you paste that command, stop. Check the URL. Verify the source. Because in 2026, the most dangerous thing a developer can do is think they already know it all.

Source: https://gbhackers.com/seo-poisoning-gemini-cli-claude-installers/

If this article made you think (or just made you angry), share it with that colleague who is always in a rush. And if you want to stay updated on how hackers are trying to steal the next big thing in AI, subscribe to the newsletter. Better safe than sorry.

Subscribe now

Glossary for anyone opening this post like it’s a box of weird words

• SEO Poisoning: A criminal technique that manipulates search engine results to push malicious sites to the top of the list, making them look legitimate or more reliable than the official ones.

• Infostealer: A type of malware specifically designed to steal sensitive data (passwords, cookies, files) from the victim’s computer and send them to the criminals.

• Fileless Malware: An attack that does not save infected files onto the victim’s hard drive, but executes the malicious code directly in the computer’s RAM, making it much harder for traditional antivirus software to detect.

• PowerShell: A powerful automation and system management tool included in Windows, which criminals frequently use because it can execute complex, hidden commands.

• ETW (Event Tracing for Windows): A Windows system that logs system activities for debugging and monitoring. Malware disables it to avoid leaving traces of its actions.

• AMSI (Antimalware Scan Interface): A Windows interface that allows programs to send scripts and files to antivirus software for real-time scanning. Disabling it means preventing the antivirus from seeing the malicious code.

• Bulletproof Hosting: Hosting services that ignore requests to take down illegal or malicious content, offering criminals a safe haven to host their command-and-control servers.

• MFA (Multi-Factor Authentication): A security method that requires two or more proofs of identity (e.g., password + code on your phone). Criminals bypass this system by stealing active session cookies.

News n°2

AI Found 10,000 Security Flaws. The Problem Is That No One Can Keep Up to Patch Them

The news comes straight from the Anthropic lab and it hits hard: their new model, Claude Mythos, has discovered over ten thousand critical security flaws—the kind known in the industry as “zero-days”—in software used by all of us. It looks like a triumph of science, the moment when artificial intelligence finally saves the day for global IT systems. Too bad reality is a bit more bitter. Our new digital superhero has ferreted out bugs at such a frantic pace that it has brought the entire security industry to its knees, revealing it’s incapable of reacting with the same speed.

Imagine having a fire hose firing bullets at a sand wall: the wall (or rather, the developers’ patching pipeline) crumbles before anyone can even figure out where the attack is coming from. Mythos analyzed more than a thousand open-source projects and turned over nearly twenty-three thousand reports. The beauty of the whole thing? The error rate is negligible, almost perfect. When it said “there’s a hole here,” there actually was one. The problem is no longer finding the vulnerabilities, because the AI finds them all, even those left there twenty years ago and forgotten by everyone. The real issue, the one that makes you laugh to keep from crying, is that the humans supposed to fix them are completely underwater.

The numbers are mind-boggling. Out of nearly two thousand cases verified by outside experts, only ninety-seven have actually been patched in the source code. Ninety-eight security advisories published. The rest? A pile of wide-open holes, waiting for someone, somewhere, to get tired of reading emails and finally decide to write a line of code to lock the door. Cloudflare found two thousand, Mozilla fixed two hundred and seventy-one flaws in a single browser version, but it’s just a drop in the ocean of bugs that Mythos keeps spitting out. Some developers, in an act of desperate honesty, have asked Anthropic to slow down. Yes, you heard that right: they are asking their most powerful tool to “pump the brakes” because they can’t keep up. It’s like a mechanic asking the car to slow down because they can’t change the tires fast enough.

There are cases that make you think, in the sense that you wonder how they could even still exist. A flaw in a cryptographic library used by billions of devices, capable of forging bank certificates, was found by the AI and patched. Another one, twenty-seven years old in the OpenBSD kernel, was uncovered by an algorithm that never slept. The AI didn’t just beat humans in speed, but in the ability to see what had been invisible for decades. Now Anthropic has launched a “Security” version for enterprises, promising to help close the gaps, but the question remains: is it possible that the only solution to an AI that finds too many bugs is another AI that fixes them? Maybe that’s the future. A war between robots, while we humans just stand by and watch, a little surprised and a little spooked.

Source: https://gbhackers.com/claude-mythos-preview-0-day/

Glossary for Outsiders

• Zero-Day (0-Day): A security hole in a software that the vendor is unaware of (or hasn’t fixed yet). It’s called that because the creator has “zero days” to fix the issue before it gets exploited. It’s like finding out your front door can be opened without a key, but only after someone has already walked in.

• Patch: The “band-aid” developers apply to software to close a security hole. Think of it like the patch you put on a flat bike tire, but for code. If you don’t do it, the software stays vulnerable.

• Open Source: Software whose code is public and can be modified by anyone. It’s like if the Coca-Cola recipe were public and everyone could tweak it (or ruin it). This allows more eyes to check it, but also more people to find bugs.

• CVE (Common Vulnerabilities and Exposures): A unique identification system for security flaws. Every hole gets a code like “CVE-2026-5194,” acting like a license plate for cybercriminals. It keeps things clear when thousands of them are being discussed.

• False Positive: When a security system says “danger!” but there isn’t actually any. It’s like a burglar alarm going off because a cat walked by. In this case, Anthropic’s AI had very few false positives, making it highly reliable.

News n°3

While You’re Swapping Memes on Telegram, They’re Breaking Into Your Network

It looks like 2025 was the year when Russian state-sponsored hacker groups decided that patience was a virtue they could do without. While the rest of us were busy wondering if our coffee was hot enough, they registered nearly 6,000 cyber incidents in Ukraine alone, a 37% jump compared to the previous year. We aren’t talking about boring little data thefts here; this is a well-oiled digital warfare campaign directed squarely at governments, defense, and the very power grid infrastructure that keeps our lights on while we decide to complain on social media.

The method is almost boring in its efficiency. They don’t need brilliant, incomprehensible mathematical code; all it takes is for you to leave your RDP exposed or to use a VPN with a flaw that should have been patched years ago. They exploit our habits: we trust whatever pops up on Signal, WhatsApp, or Telegram, and that’s where their stuff slips in. It’s not magic, it’s just social engineering mixed with a total lack of digital hygiene on the part of the victims. If they send you an SVG file or an LNK shortcut link on Dropbox and you click it, well, welcome to the club. They now have the keys to your kingdom.

And the most cynical part? They often don’t even need the classic phishing emails full of grammar mistakes anymore. They buy ready-made credentials on darknet forums and slip right into the infrastructure, completely skipping the “courtship” phase. Once inside, they use your very own tools against you: PowerShell, certutil, the whole system administration toolkit you’re supposed to use for work, weaponized to hide in the background noise. If you think old Microsoft Office vulnerabilities are ancient history, think again. They’re still right there, untouched, just waiting for someone to click an attachment.

The solution? It seems obvious, but it’s like saying that to stop overeating, you just need to stop eating. You’re supposed to use multi-factor authentication, adopt Zero Trust architectures, take patch management seriously (yes, even for the old stuff), and above all, stop blindly trusting anyone who messages you. But as long as you think a security update is a nuisance, you’ll always be the first weak link in the chain. And as for them, well, they’re already counting their cash.

Source: https://cybersecuritynews.com/russian-threat-groups-use-rdp-vpn-supply-chain-attacks/

Glossary for Anyone Who Isn’t a Hacker (and Hopes They Never Have to Become One)

• RDP (Remote Desktop Protocol): A protocol that lets you control a computer remotely. If exposed to the internet without proper protection, it’s like leaving your front door wide open with the key under the welcome mat.

• VPN (Virtual Private Network): A tool to create a secure connection over the internet. If it has known (and unpatched) vulnerabilities, it becomes a direct tunnel for attackers right into your internal network.

• Supply Chain Attack: An attack targeting vendors. Instead of hitting the final target directly, hackers infect a supplier or software that the target uses, gaining “rightful” entry into their systems.

• Social Engineering: The art of manipulating people into handing over information or access. You don’t need to be a programming genius if you’re good at lying or triggering panic.

• Wiper Malware: A type of malware designed not to steal data, but to permanently delete it, rendering systems completely unusable. Pure destruction.

• Living off the Land: A technique that involves using tools already present in the system (like PowerShell) to carry out malicious actions, making it hard for security systems to tell the difference between legitimate activity and an attack.

• Zero Trust: A security model that doesn’t trust anyone, not even those who are already inside the network. Every single access request must be continuously verified.

• IoC (Indicator of Compromise): Digital breadcrumbs (IP addresses, file hashes, domains) left behind by hackers that allow analysts to realize an attack is underway or has already happened.

News n°4

The Great Middle Eastern Circus: How Hackers Transformed Telecoms into a Playground

There are people who buy a villa to live in it, and then there are those who turn the phone network of an entire continent into a parking lot for cybercriminals. It sounds like a punchline from a stand-up comedy routine, but it’s the grim reality described in a new report by Hunt.io: in the Middle East, hackers have decided it’s much more convenient to use the cables of telecom giants than to build their own.

Imagine opening a pizzeria only to find out that your neighbor, instead of cooking, has decided to use your oven to hide illegal weapons. That’s not far from what’s actually happening. In just three months, over 1,350 command-and-control (C2) servers were identified across 14 countries in the region. The data point that makes you laugh to keep from crying isn’t so much the sheer number, but the concentration: Saudi Telecom Company (STC), all by itself, hosts 981 of these servers, meaning nearly 73% of all regional malicious activity. It’s as if a single restaurant played host to 73% of all the criminals in an entire city. And the most cynical part? It’s not like the company invited these guests over for dinner. It’s almost certain they simply picked the lock to the house and sat down at the table without asking for permission, turning compromised customer devices into a military base of operations.

It’s not just a numbers game; it’s a bizarre cocktail of expertise. On the very same networks that are supposed to guarantee your connection for video calls, you have IoT exploitation botnets running alongside penny-ante phishing kits and cyber-espionage tools so sophisticated they’d make nation-states jealous. It’s as if a bike thief and an international spy agency lived in the exact same apartment building. The fact that the provider Türk Telekom hosts malware from six different families, or that in Iraq, a provider called Regxa has a “protection” score of practically zero (meaning they ignore abuse reports), is proof that cybersecurity in this region has become a role-playing game where the hackers wrote the rules.

The specific campaigns are a non-stop parade of paradoxes. In Syria, the Phorpiex botnet uses phone lines to deploy crypto-miners that hijack your PC’s processing power to line the pockets of someone you’ll never meet. In Iraq, spies from the Eagle Werewolf cluster pretend to be drone delivery services or Starlink registrations to infect devices. In Saudi Arabia, they exploit a software flaw to render users’ computers useless, while in Iran, the RondoDox botnet launches thousands of attacks a day against anyone with a powered-on device. The takeaway, if one were even needed, is that instead of chasing down individual IP addresses like rats in a maze, defenders should stop worrying about who is shooting and start looking at who gave them the gun. Monitoring providers and networks, rather than individual threat actors, is the only way to understand the game without getting fooled by the tricks.

Source: https://cybersecuritynews.com/hackers-abuse-middle-east-telecom-networks/

Glossary for Survivors (or at least for anyone who wants to make sense of this)

• Command-and-Control (C2): Imagine a command room where a criminal gives orders to their robots. It’s the central server that controls all other infected devices, giving them instructions on what to do (steal data, send spam, etc.).

• Botnet: An army of “zombie” computers controlled remotely by a hacker. Each infected computer is a soldier carrying out orders without ever knowing it.

• ASN (Autonomous System Number): The identity code of an internet network, similar to a phone number for a telecommunications company. It helps track down who owns the network.

• Phishing: The deception technique where hackers pose as trustworthy institutions (banks, postal services) to trick users into revealing passwords or sensitive data, usually through fake emails or websites.

If you want to stay updated without having to read 50-page technical reports, subscribe to the newsletter so you don’t miss the next stories on how the digital world is making fools out of us.

Subscribe now

2.1 Basic Formatting and Data Formats

All primary visual management tools are located in the quick access toolbar, situated immediately below the main menu.

From this bar, you can adjust the typographic appearance: by selecting a cell or a range, you can change the font, adjust its size, and apply styles such as bold, italics, or strikethrough. It is also possible to alter the text color and the fill color (background) of the cell to create visual hierarchies.

Text orientation can be managed on two axes:

• Horizontal alignment: Allows you to position content to the left, center, or right. By default, Google Sheets aligns text to the left and numbers to the right, but this setting can be overridden.

• Vertical alignment: Allows you to position content at the top, middle, or bottom of the cell space. You can also set Text wrapping so that long content isn’t visually cut off, but instead flows onto multiple lines within the same cell, expanding its height.

AVOID Strictly avoid using “creative” fonts, elaborate scripts, or handwriting-style fonts (e.g., fonts that mimic calligraphy). Using non-standard characters drastically degrades the spreadsheet’s readability, making it look unprofessional. Stick to clean, crisp fonts like Arial or Verdana.

Beyond aesthetics, it is mandatory to instruct the software on the type of data entered by assigning a Format. Data formatting alters how the system processes it mathematically:

• Currency Format: Automatically adds the currency symbol (e.g., $) and formats values with standard decimals.

• Percent Format: Transforms a decimal number into its mathematical percentage representation (e.g., 0.1 becomes 10%).

• Plain Text Format: Forces the system to treat a number (such as a barcode or an Order ID containing commas) as a text string, preventing it from performing accidental calculations or formatting it incorrectly. For numbers and currencies, the toolbar offers quick buttons to increase or decrease the displayed decimal places.

2.2 Managing Rows and Columns

Data grids require constant dimensional adjustments to accommodate text of varying lengths.

• Resizing: The manual method involves placing the mouse cursor on the right edge of a column letter (or the bottom edge of a row number): the cursor will change shape, allowing you to click and drag to widen or narrow the space. The automatic method (Auto-fit) is much more efficient: by double-clicking the border between two column letters, the software will instantly adjust the width to perfectly fit the longest data entry in that specific column.

PRO TIP To automatically resize the entire worksheet in one go, click the empty rectangle at the top left (exactly between the letter A and the number 1) to select the entire grid. Then, double-click the border of any column: all columns will simultaneously adapt to their respective contents.

• Inserting and Removing: To add new structural space, right-click on the column letter or the reference row number. From the context menu, you can select specific commands like “Insert 1 column left/right“ or “Insert 1 row above/below.” The same right-click procedure allows you to select “Delete column“ or “Delete row“ to remove entire sections of the grid.

2.3 Advanced Visual Tools

The interface offers three advanced tools to optimize the usability of extensive databases without wasting time on repetitive manual operations.

• Format Painter: Represented by the paint roller icon, this tool allows you to replicate a complex set of styles from one cell to another. Procedure: Select the cell with the desired formatting (e.g., yellow background, bold text, centered), click the Format Painter icon, and finally click the destination cell (or range of cells). The system will apply the entire aesthetic combination with a single click.

• Alternating Colors: To improve visual tracking across very long rows, Google Sheets allows you to automate the application of colored backgrounds to every other row. By selecting the data range and accessing the menu Format > Alternating colors, a side panel opens. Here, you can define a style for the header row and two colors that will automatically alternate for all subsequent rows, drastically improving the database’s readability.

• Freeze Rows/Columns: When scrolling through a database containing hundreds of records, column headers tend to disappear off the top, making it impossible to know what the data refers to. From the menu View > Freeze, you can “lock” a specific number of rows or columns. By anchoring Row 1, for example, the header will remain fixed and always visible at the top of the screen regardless of how far down you scroll. A quick manual method to achieve this is to place the cursor on the thick gray line in the top-left corner of the grid and drag it down (for rows) or to the right (for columns).

News n°1

When Renting Servers Becomes More Profitable Than Winning the Race

Elon Musk, the visionary who promised to revolutionize AI as well, finds himself facing a rather embarrassing problem: his Grok is looking like a runner huffing and puffing at the back of the pack while the other sprinters zoom past him. And the solution he’s come up with? Renting out his supercomputers to one of the very rivals who is currently outclassing him.

The deal signed in May 2026 is a real head-scratcher: Anthropic, the company behind Claude AI, has secured exclusive use of Colossus 1, one of Musk’s most powerful data centers near Memphis. In exchange, several billion dollars a year are rolling in, coming in clutch just as SpaceX prepares for its IPO. In short, if you can’t beat ‘em, at least charge ‘em rent.

The numbers tell a story that no optimistic press release can mask. Grok app downloads plummeted from over 20 million in January to just over 8 million in April. The percentage of Americans willing to pay for Grok is effectively stuck at 0.174%, a figure that would make anyone cry when compared to the 6% paying for ChatGPT.

In the enterprise world, where the real money is made, Grok is even more marginal. Only 7% of companies surveyed in March 2026 reported using it and intending to continue, a negligible growth from 4% the previous year. Meanwhile, Claude has skyrocketed from 21% to 48%, and Gemini from 27% to 40%. Clearly, integrating a chatbot with X isn’t enough when the competition is leaving you in the dust on quality.

And then there’s that “cherry on top” that caused downloads to spike in January: a feature that allowed users to “virtually undress” people in photos. When it emerged that it was being used on images of minors, regulators and lawmakers immediately went on high alert. Perhaps not exactly the kind of virality you were looking for when you launched your chatbot at the end of 2023.

So here we are: Musk raking in billions by renting his servers to Anthropic while Grok struggles to keep up. It makes you wonder if, in the AI arms race, he hasn’t simply decided to switch tracks—moving from AI developer to digital landlord. Less glorious, sure, but at least it pays the bills.

Source: https://www.wsj.com/tech/ai/anthropic-spacex-ai-deal-elon-musk-f86ea369?st=XUQnP7&reflink=desktopwebshare_permalink

Minimal Glossary

• IPO (Initial Public Offering): The first sale of a company’s stock to the public, allowing everyday investors to buy shares. In SpaceX’s case, it would mean finally being able to purchase stakes in Musk’s space company.

• Enterprise: A term referring to the market of large corporations and organizations, as opposed to individual consumers. Enterprise adoption is crucial because businesses pay significantly more than individual users.

News n°2

In the wonderful world where anything containing the word “AI” is automatically deemed trustworthy, someone thought it would be a great idea to turn the hype for DeepSeek TUI into a convenient vehicle for distributing malware to distracted developers and techies. DeepSeek TUI is a legitimate terminal tool for interacting with DeepSeek models; too bad that, surrounding the real project, fake GitHub repositories have blossomed, mimicking its graphics, structure, and naming with the sole mission of infecting anyone who downloads the wrong “setup.”

The scheme is the same old story, but refined with care: masterfully cloned repositories, a Releases section full of 7z archives that look like normal installers, reassuring names, and a perfect context, the release of DeepSeek v4 and viral posts in Chinese tech communities, to lower the guard of those more eager to play with the latest toy than to read two extra lines of text. Behind this theater, researchers at QiAnXin recognized virtually the same tricks, infrastructure, and domains previously seen in a campaign called OpenClaw, a sign that the actor is the same as always, just switching masks while the audience applauds.

To make things more “fun,” the campaign doesn’t stop at DeepSeek TUI but abuses every possible buzzword: fake installers for tools dubbed Claude, Grok, WormGPT, KawaiiGPT, fraudGPT, and the like, all linked by the same PDB “ClawCode.pdb” and the same malware family written in Rust. The main file, DeepSeek-TUI_x64.exe, checks if it’s being watched before even getting its hands dirty: if it sniffs out a sandbox, VM, or analysis tools, it feigns innocence with a “your system does not meet minimum requirements” message and quietly shuts down, because even malware has learned to do OPSEC better than many IT teams.

If, however, it realizes it’s on a real machine, the show begins: using an obfuscated PowerShell script, it disables key Windows Defender protections, adds folder exclusions, turns off behavior monitoring, and opens firewall ports (57001, 57002, 56001) to pave the way for subsequent payloads. At that point, it contacts Pastebin and snippet.host to retrieve second-stage components hosted on Azure, using links and passwords saved there as if they were harmless text notes; it’s hard to imagine a more creative use of “developer services.”

The subsequent modules are specialized: OneSync.exe and WinHealhCare.exe manage installation, scheduled tasks, and reporting via Telegram, while onedrive_sync.exe ensures the malware survives reboots by stuffing keys into the Registry’s Run folder. The resident heart, svc_service.exe, uses direct NT calls to inject threads and load .NET assemblies into memory only, minimizing the disk footprint and maximizing headaches for defense teams. To complete the picture, the autodate.exe loader disguises itself as a service manager while actually injecting payloads into memory and talking to command-and-control servers on domains like mikolirentryifosttry.info and zkevopenanu.cfd, because calling itself “bad-malware-server.evil” would have been too honest.

Persistence is redundant to an almost paranoid degree: scheduled tasks, Run keys, Winlogon hooks, startup shortcuts; once a system is compromised, a simple “scan” isn’t enough to clean it up and go back happily to your coding session. And for those wishing to analyze the specimens, the article lists a long string of MD5 hashes associated with executables sporting reassuring names like DeepSeek-TUI_x64.exe, GrokCLI_x64.exe, WormGPT_x64.exe, KawaiiGPT_x64.exe, along with all second-stage components and URLs used for staging and passwords.

The moral, of course, comes at the end, but it was the real news from the start: developers should stop treating GitHub like an official app store with divine moderation and start acting as if anyone could upload anything (because that’s exactly the case). Checking account age, commit history, actual contributors, and project reputation should be as trivial as typing git clone, yet we continue to see people installing “deepseek-v4-pro_x64.exe” caught from the first trendy repo that passes by. Security tools that monitor memory injections and creative PowerShell use can help, but if the first line of defense is a trusting click on the “Download” button, the rest is just post-disaster mitigation.

Source: https://cybersecuritynews.com/hackers-use-fake-deepseek-tui-github-repositories/

Glossary for the Uninitiated

• DeepSeek TUI: A terminal interface that allows you to use DeepSeek language models directly from the command line—no shiny GUI, but plenty of flexibility.

• GitHub Repository: A space (public or private) on GitHub where a software project’s code, files, and history live; anyone can create one, including attackers.

• 7z (Compressed Archive): A file compression format similar to zip and rar; in this case, it’s used to bundle malware disguised as a legitimate installer.

• Malware: Software designed to damage, spy on, or take control of a system; here it is distributed by posing as popular AI tools.

• Sandbox: An isolated environment where analysts and antivirus software run suspicious programs safely; many malwares try to detect if they are in a sandbox and, if so, “play nice.”

• PowerShell: A very powerful Windows shell and scripting language; extremely useful for admins, and equally useful for attackers wanting to automate malicious operations.

• Windows Defender: The security suite built into Windows that includes antivirus and various protections; in the case described, it is deliberately weakened by the malware to give it free rein.

• Second-stage payload: An attack component downloaded after the initial infection; it usually handles more advanced functions like persistence, data theft, or remote control.

• Persistence: A set of techniques that allow malware to remain active even after reboots or logouts, using things like scheduled tasks and registry keys.

• C2 (Command and Control): A server controlled by the attacker that the malware connects to for instructions and to send data; it’s the operation’s “mission control.”

• Indicators of Compromise (IoC): Technical signs of an infection, such as file hashes, domains, and URLs used by the attacker; they are used to detect and block threats in security systems.

News n°3

When Microsoft AI Becomes an Open Book (for Hackers)

Ah, Microsoft. Always ready to surprise us with new vulnerabilities just when we thought entrusting our entire corporate infrastructure to an AI assistant was a brilliant idea. On May 7, 2026, Redmond confessed to the existence of three security flaws in Microsoft 365 Copilot and its Edge incarnation—all capable of leaking sensitive information like it was office gossip.

The three vulnerabilities share the same severity score (7.5 out of 10) and the same worrying trait: an attacker doesn’t need special credentials or a way to trick some poor soul into clicking a suspicious link. You just need to know the right trick and you’re in. The first flaw, labeled CVE-2026-26129, allows for the manipulation of how Copilot processes inputs due to shoddy handling of special elements. The second, CVE-2026-26164, is a classic injection problem where Copilot’s output can be tainted to deceive downstream components. The third, CVE-2026-33111, specifically hits the Edge version and even allows for command injection.

The real problem is that Microsoft 365 Copilot isn’t some isolated toy: it has its fingers in every pie, from emails and Word documents to Teams chats and SharePoint files. Imagine a corporate assistant that knows all the company secrets—from financial statements to future strategies—and can suddenly be convinced to spill the beans through cleverly crafted prompts. Not exactly the level of confidentiality one expects from a tool sold as a revolutionary game-changer for business productivity.

The silver lining, if we want to find one, is that Microsoft has already fixed everything on the server side without customers having to lift a finger. No patches to install, no reboots, no panic. That’s the beauty of fully managed cloud services: holes get plugged silently and we can all go back to pretending nothing ever happened. Until the next vulnerability disclosure, of course.

Source: https://gbhackers.com/microsoft-365-copilot-flaws/

Glossary for Newbies

• CWE (Common Weakness Enumeration): A catalog that classifies types of software weaknesses—basically a manual of every way a programmer can mess up.

• Prompt injection: An attack technique that exploits how language models interpret instructions, tricking the AI into doing things it shouldn’t by manipulating the input provided.

• Backend: The part of an online service that runs on the company’s servers; invisible to the end user but responsible for the actual functioning of the system.

• Network Attack Vector: Indicates that the vulnerability can be exploited remotely over the Internet, without needing physical access to the target system.

News n°4

When Google Ads Becomes an (Unintentional) Accomplice to Cybercrime

If you thought paying Google to appear at the top of search results was just a matter of aggressive marketing, get ready to discover that cybercriminals have turned this practice into a fine art. The new frontier of malvertising targets macOS users with a strategy that is as simple as it is effective: placing fake ads that mimic legitimate AI applications, specifically Claude.

The kicker is that we aren’t talking about amateur hour operations run by some script kiddie in their parent’s garage. These attackers have realized that to look credible, they need to leverage the infrastructure of tech giants: Google Sites, Framer, and even authentic shared chats from claude.ai become unwitting hosts for malicious pages. Why bother setting up shady servers when you can hide behind the reliability of platforms that no corporate filter would ever dare to block?

The preferred tactic is called “ClickFix,” and it’s a masterpiece of social engineering. Imagine searching for Claude AI, clicking on what looks like the official site, and finding yourself in front of an interface so well-made it looks like the real deal. However, when you attempt to download it, dialog boxes appear, convincing you to run terminal commands or install compromised software. At that point, MacSync, the malware protagonist of this story, begins its harvest: browser credentials, crypto wallet data, session tokens. It’s basically an all-you-can-eat buffet of your personal information.

Researchers Berk Albayrak and g0njxa have documented malicious URLs hosted on Google Sites with reassuring names like “cloud-version-08” or “claud-version-0505.” Notice the typo? It’s not exactly subtle, but evidently, it’s enough to fool someone in a hurry. And when a domain gets burned, no problem: they just create another, move the operation to a different platform, and the cycle starts all over again.

The solution proposed by experts is the usual refrain: block indicators of compromise, monitor anomalies in system logs, and avoid clicking on sponsored ads. These are sound pieces of advice that nevertheless ignore reality: the vast majority of users cannot distinguish an ad from an organic result and will continue to trust the fact that “if it’s on Google, it must be safe.” Spoiler alert: it’s not.

Source: https://cybersecuritynews.com/macos-malware-leverages-google-ads/

Cyber Jungle Survival Glossary

• Malvertising: The practice of using legitimate online advertising (like Google Ads) to distribute malware or redirect users to malicious sites.

• ClickFix: A social engineering technique that manipulates users through deceptive dialog boxes, tricking them into performing harmful actions under the guise of fixing a technical issue.

• Information stealer: A type of malware specifically designed to steal sensitive data from an infected device, such as passwords, banking info, or login credentials.

• Indicators of Compromise (IoC): Technical elements (URLs, file hashes, IP addresses) that signal the presence of malicious activity in a system or network.

• Session token: Temporary credentials that allow you to stay logged into a web service without repeatedly entering your username and password; if stolen, they allow attackers to impersonate the victim.

• Telemetry: The automatic collection and monitoring of data regarding a system’s behavior and activities, used to identify anomalies or threats.

News n°5

OpenAI Is Back to Explaining Why “Nothing Happened”

OpenAI has once again found itself explaining why “nothing happened”, except for the tiny detail that hackers managed to steal some internal credentials by piggybacking on an open-source library used by the company.

According to the official version, two employees had their devices compromised after attackers took control of TanStack, an open-source project for web app development. The hackers published 84 infected versions of the software in a six-minute window, which were only detected about twenty minutes later. The malicious code was designed to harvest credentials from the computers where it was installed and to self-spread to other systems, because automation is great for everyone, even those writing malware.

Following an internal investigation, OpenAI is repeating the mantra: no user data was touched, no production systems were compromised, no intellectual property was stolen, and no software installations are at risk. The problem, the company assures, is limited to the devices of the two employees and a “limited subset” of internal code repositories they had access to, from which only a small amount of “credential materials” were allegedly taken. It’s a shame, however, that those repositories also contained digital certificates used to sign products. This is why the company is now forced to rotate them “as a precaution,” with the lovely consequence that macOS users will have to update their apps.

This all fits into the now-recurring trend of supply chain attacks: instead of going directly after the big fish, hackers target widely used open-source libraries and tools, slipping malware into updates that look completely normal. It’s the same logic recently seen with the hijacking of Axios, a library used by millions of developers, attributed to North Korean hackers, and the accusations against a Chinese group for a similar attack on Daemon Tools software, which runs on thousands of Windows machines. In the TanStack case, it’s not yet clear who is behind it, though the article cites a group called TeamPCP as a precedent, a group well-known enough to have been targeted by other hackers themselves, because in this industry, the victim hierarchy is rather fluid.

In the end, it’s the same old picture: a modern software infrastructure built on layers of open-source components that everyone takes for granted until someone compromises them, and big companies reassuring everyone that everything is fine, aside from a few “technical details” like rotating certificates and stolen internal credentials.

Source: https://techcrunch.com/2026/05/14/openai-says-hackers-stole-some-data-after-latest-code-security-issue/

Minimal Glossary for the Tech-Illiterate

• Open source: Software whose code is public and modifiable by anyone, often maintained by developer communities; it’s used en masse even by big companies because it’s convenient and free.

• Supply chain attack: A technique where an “upstream” link is hit (like a library or development tool) instead of launching the attack directly against a single company, allowing the impact to trickle down to many targets at once.

• TanStack: An open-source project providing libraries to help developers build web applications; it became the vehicle for the attack after hackers published infected versions of its packages.

• Source code repository: An archive where the source code of a software project is saved, often hosted on platforms like GitHub; it contains code and, sometimes, secrets that shouldn’t be there, like credentials or certificates.

• Credentials: Data that allows access to systems or services, such as usernames, passwords, access tokens, or API keys; if stolen, someone can impersonate legitimate users or services.

• Digital certificates: Cryptographic files used to guarantee the identity of software or a service; for example, they sign an application so the operating system recognizes it as authentic.

• Certificate rotation: A procedure where potentially exposed certificates are invalidated and new ones are issued, often forcing users to update applications or configurations.

Usually, the scene goes like this: you stop writing, start scrolling frantically up with your mouse, hunt for the exact spot, lose the inspiration for what you were currently working on, and end up staring at a cursor that’s blinking mockingly at you. Frustrating, right? Well, just so you know, I’ve been there so many times that my mouse was about to file for divorce. Then I discovered that Google Docs lets you jot everything down without ever taking your hands off the keyboard.

The Trick to Staying Focused

The secret to not losing your mind (and not losing your ideas along the way) lies in learning to delegate your thoughts to little virtual assistants: comments and notes. Instead of rewriting everything every time you have a doubt, you can simply “pin” the thought and keep racing toward the end of the document.

Managing a long document without these shortcuts makes everything extremely chaotic.

Let’s Get to Work (No Mouse Needed!)

Here’s how to become a true word-processing ninja. Open a Google Doc and try these two magic tricks:

#1 The Instant Comment

If you want to leave a reminder for yourself (or a colleague) about a specific word:

• Select the word or place your cursor at the end of the sentence.

• Press Ctrl + Alt + M at the same time.

• A little bubble will magically appear on the right. Type your note and hit Enter.

This feature is useful because it allows you to keep reading the main text without interruption; the note stays “out of the way” until you’re ready to deal with it.

#2 The Footnote

If you want to get fancy and insert a citation or an academic explanation at the bottom of the page:

• Place the cursor after the word you want to explain.

• Press Ctrl + Alt + F.

• The cursor will instantly fly to the bottom of the page, create a tiny little number, and let you type in your source.

Once you’re done writing the note, just hit the up arrow or click back into the text to return exactly to where you left off.

Today’s Takeaway

Today we learned that while the mouse is a great tool, for pure writing, it’s our worst enemy. Using Ctrl + Alt + M for comments and Ctrl + Alt + F for footnotes lets you keep your brain focused on the goal: finishing that blessed page without distractions.

What about you? Are you the “write it all in one go and edit later” type, or do you stop every two minutes to fix every single comma?

Quick Glossary to Keep You on Track

Shortcut: A key combination that executes a command that would normally take several mouse clicks.

Flow: That mental state where you’re so focused on an activity that time seems to fly and everything comes easily.

Interface (UI): Everything you see on the screen and interact with (buttons, menus, windows).

Subscribe now

To start working with Google Sheets, the fastest and most efficient approach doesn’t require navigating through complex menus. By typing the shortcut sheet.new or sheets.new directly into your browser’s address bar, the system instantly generates a new blank document ready for use.

The user interface is divided into three primary operating sections:

• The title bar: Located at the top left, it allows you to rename the file by clicking on the default text “Untitled spreadsheet.” Saving happens automatically to Google Drive with every change.

• The drop-down menus: The top bar (File, Edit, View, Insert, Format, Data, etc.) contains the entirety of the software’s commands and advanced features.

• The quick access toolbar: Positioned immediately below the menus, it displays icons for the most frequently used functions, such as text formatting, borders, alignment, and filters.

PRO TIP Always rename your file as your first step. Entering large amounts of data into “Untitled” documents makes them extremely difficult to retrieve later within the Google Drive cloud architecture.

1.2 The Anatomy of a Spreadsheet

To operate effectively, it is mandatory to use the correct technical terminology. The spreadsheet is based on a grid whose coordinates define the exact position of the data.

• Columns: These are the vertical elements of the grid, identified by letters (A, B, C, etc.) located in the top header.

• Rows: These are the horizontal elements, identified by numbers (1, 2, 3, etc.) located on the left margin.

• Cells: The exact intersection between a column and a row creates a cell. Standard nomenclature always lists the column letter first and then the row number (e.g., A1, D6, I10). The identifier of the active cell is always visible in the “Name box” located at the top left.

• Ranges: A contiguous group of selected cells. The technical syntax for defining a range requires indicating the top-left cell and the bottom-right cell, separated by a colon (:), which functions as “through.” A formal example is D2:H13.

• Sheets and Workbooks: The entire file you create is called a “Workbook.” This workbook can contain multiple “Sheets“ within it, which can be managed using the tabs located at the bottom left of the interface.

1.3 Data Entry

The golden rule for data entry is the principle of “select to affect“: to make any change, the target cell must be selected first.

It is crucial to understand the difference between two seemingly similar actions that yield opposite results:

• Clicking ON a cell (Single click): The cell is selected. Any data typed on the keyboard will entirely overwrite the existing content.

• Clicking IN a cell (Double click): You enter “edit mode” inside the cell. A blinking cursor appears, allowing you to alter specific parts of the text without deleting the entire content.

To navigate the grid during continuous data entry, using a mouse is inefficient. The correct protocol involves using the keyboard:

• The Enter key confirms the entry and moves the selection to the cell below.

• The Shift+Enter combination moves the selection to the cell above.

• The Tab key moves the selection to the cell to the right.

• The Shift+Tab combination moves the selection to the cell to the left.

AVOID Avoid taking your hands off the keyboard to use the mouse after every single data entry. Using the arrow keys or the Tab/Enter commands drastically reduces database compilation time.

1.4 Basic Entry Automation

Google Sheets features built-in tools to speed up the entry of repetitive or sequential data.

The Fill Handle

When you select a cell or a range, a small blue square or dot appears in the bottom-right corner of the selection border. By clicking and dragging this point:

• If you select a single cell containing text or a number, the software will replicate the same content in all the dragged cells.

• If you highlight two cells showing a pattern (e.g., the numbers 1 and 2), dragging the handle will cause the system to extend the numerical series sequentially (3, 4, 5...).
  The system automatically recognizes complex patterns like dates, days of the week, or formats like “Week 1,” extending them correctly (e.g., Week 2, Week 3) both vertically and horizontally.

Smart Fill

For more complex text patterns, Google Sheets integrates a Smart Fill function. If, for example, you have a column with First and Last Names and you begin manually typing email addresses derived from those names in an adjacent column (e.g., first.last@email.com), the system will detect the pattern after a few rows. An automatic suggestion will appear; by confirming it (via the checkmark or the Ctrl+Shift+Y shortcut), the software’s intelligence will instantly populate the rest of the column, performing the operation for you.

Subscribe now

News N°1

The Crypto Highway of the Ayatollahs: The Nobitex Case

In Iran, even the digital revolution remains a family affair: the offspring of one of the country’s most influential dynasties created Nobitex, the nation’s leading crypto exchange, transforming it into a convenient parallel financial highway. It is used by both ordinary citizens and sanctioned institutions like the central bank and the Revolutionary Guard (IRGC), all while the West has yet to find the time to actually sanction the platform itself.

On paper, Nobitex is a classic startup brochure story: three young graduates from Sharif University (the Iranian MIT) who, in 2018, launched an exchange to help fellow citizens crushed by inflation, a free-falling rial, and bans from international banks to move their savings into cryptocurrencies. In practice, behind the two founding brothers, Ali and Mohammad, stands not just another team of nerds, but the Kharrazi clan: an ayatollah grandfather who is a member of the Assembly of Experts, a hardline preacher father who founded a political group called Hezbollah in Iran, and relatives married into the inner circles of all three Supreme Leaders of the Islamic Republic, including the current Mojtaba Khamenei. However, the young men present themselves to the public using a different last name, Aghamir, and for years, even friends and colleagues had no idea they were hanging out with scions of the religious power’s inner circle. The family tree was eventually exposed by corporate documents, registries of a religious foundation, and even the email used to register the Nobitex domain, which, wouldn’t you know it, contains the surname “Kharrazi.”

Within a few years, Nobitex became the main lung of the Iranian crypto market, boasting 11 million declared users, over 10% of the population, and handling about 70% of the country’s digital transactions. For the average citizen, unable to open accounts on giants like Binance, the exchange is the gateway to a pseudo-global world that is poorly regulated and far more flexible than traditional banks. For the State, particularly the central bank and the Revolutionary Guard, it is a tool to move capital in tokenized dollars and other cryptocurrencies under the radar of Western sanctions, exploiting the usual ambiguity: within the same infrastructure, millions of micro-transactions from ordinary people are mixed with cash flows from institutions on U.S. and allied blacklists. Analyses from firms like Elliptic, Chainalysis, and Crystal Intelligence indicate that between 2024 and 2025, the central bank’s network purchased over $500 million in crypto, of which approximately $347 million allegedly flowed into Nobitex in the first six months of 2025 alone. These figures, even at the highest estimate of $366 million in “dirty” movements associated with sanctioned wallets, represent only a small percentage of the platform’s total volume of over $11 billion, but they are clearly enough to be useful to the regime.

Nobitex, for its part, sticks to the standard script: we are private and independent, with no agreements with the central bank, the IRGC, or other government bodies; any illicit funds passed through “without our knowledge” and in much smaller amounts than analysts claim. Meanwhile, in the same public documents, the company boasts of having developed cryptographic tools to “protect” users from sanction-related checks, of constantly changing wallet addresses to make tracking more difficult, and of suggesting that clients segment transactions across multiple addresses to confuse Western monitors. All this in a country where, as a former high-ranking U.S. Treasury official points out, any activity that becomes truly profitable immediately sees the State, and specifically the security apparatus, show up at the door to demand its cut. It is no surprise, then, that the platform continues to operate and grow despite IRGC visits to the offices, searches, the temporary arrest of a co-founder, and periodic banking freezes imposed by the central bank.

The group’s internal trajectory is the usual tech fairytale with a few extra surreal details. As students, the brothers lived relatively modestly, cheap cars, no flashiness, carrying around those endless surnames that amused friends but didn’t require explanation. With Nobitex’s expansion came penthouse offices in Tehran, lounge areas, video games, no hijab requirement for female employees, and staying open even during religious holidays; rather than the fintech branch of a theocracy, it looked like a coworking space in any global tech capital. Then came the 2022 protest cycle and the death of Mahsa Amini. Along with general repression, hijab enforcement and crackdowns on corporate dress codes returned through the back door, alongside increased security surveillance within the office, as State morality reclaimed its space even in the “showcase” company. Meanwhile, one of the main investors, Mohammad Bagher Nahvi, turns out to be the vice president of an airport services company, Safiran, sanctioned by the U.S. for facilitating flights that transferred drones and military personnel from Iran to Russia. When the exchange was hacked in 2025 and about $90 million in crypto ended up in wallets blocked and renamed with anti-IRGC insults, Nobitex was still able to reimburse all clients directly, confirming that even after the “damage,” the brothers and partners aren’t exactly struggling for cash.

Nobitex’s role in the sanction-evasion system emerges in an almost grotesque fashion from the online vitriol of another textbook character: Babak Zanjani, a billionaire specialized in finding holes in the sanctions regime, sentenced to death for fraud in 2016, pardoned in 2024, but still in an open war with the Iranian central bank. In a social media outburst, Zanjani published wallet addresses used in operations with the central bank, unintentionally handing external analysts a key to mapping part of the evasion architecture: at least $20 million in sanctioned funds passed through addresses later linked to Nobitex. By cross-referencing these wallets with those already identified and sanctioned by the United States and Israel, analytics firms can estimate the size of the “unclean” portion of the flows touching the exchange: it ranges from about $22 million in direct transfers from sanctioned wallets identified by Crystal Intelligence to much higher figures when considering multi-level paths, mixers, and intermediate steps. Amidst this, links also emerge to addresses associated with the Houthi rebels in Yemen, another actor aligned with Tehran and under Western sanctions, who are thus part of the same circuit. All this, of course, while Iranian authorities officially continue to operate through hawala, offshore banks, and their “shadow banking architecture,” which is now formally targeted by new U.S. sanctions that, strangely enough, still do not include Nobitex.

The picture becomes even more interesting with the war between the United States, Israel, and Iran that began in late February 2026, when the regime’s response to internal tensions was the usual reset: an almost total internet blackout for the population, with access granted only to a very narrow “whitelist” estimated at between 1% and 2% of Iranians. In this scenario, the majority of citizens cannot even read their emails, yet Nobitex, mysteriously, continues to operate and process transactions worth over $100 million, about 20% of its normal volume, in the middle of the blackout. Also according to Crystal Intelligence, at least $54 million was withdrawn from the platform during the war by this micro-group with privileged network access, and a large portion of these funds quickly disappeared abroad, converted into cash by brokers who aren’t inclined to ask questions. In parallel, the company published a reassuring message promising customers that, “despite infrastructure instability,” their money is safe and accessible, carefully avoiding any mention of the war while missiles rained down on the homes of the founders’ relatives. On the same day Nobitex was reassuring the market, a raid killed the wife and later the great-uncle of the brothers, former Foreign Minister Kamal Kharrazi, and the new Supreme Leader, himself mourning personal losses from the bombings, hurried onto TV to praise “the distinguished Kharrazi family.”

Viewed from the outside, Nobitex is the perfect product of its surrounding ecosystem: large and popular enough to be marketed as an essential service for millions of Iranians, yet connected enough to the elite and opaque enough to be useful to the regime for moving money where banks can no longer go. Senator Elizabeth Warren describes the situation as a “flashing red light,” emphasizing how adversaries are using digital assets to build an alternative financial system led away from U.S. oversight, bypassing anti-money laundering controls that, in the crypto world, are often more about marketing than substance. Meanwhile, no one has yet had the courage, or the interest, to strike the exchange directly, which holds together the savings of millions of users and the cash flows of sanctioned institutions because, as one analyst notes, it is almost impossible to fully distinguish the regime from the population when you force them through the same funnel.

Source: https://www.reuters.com/investigations/one-irans-most-powerful-families-founded-its-largest-crypto-exchange-its-used-by-2026-05-01/

Essential Glossary #1

Cryptocurrency

A digital currency based on cryptography, traded on decentralized networks and not issued by a central bank. It is often used for speculation, cross-border payments, or, when convenient, to evade financial controls.

Exchange (Trading Platform)

An online service that allows users to buy, sell, and convert cryptocurrencies—either between different digital assets or with traditional currencies. Nobitex is the dominant exchange in Iran.

Blockchain

A distributed digital ledger that stores cryptocurrency transactions in a chronological and unalterable way. It is publicly accessible, though identities are usually pseudonymous.

Stablecoin

A cryptocurrency whose value is pegged to a “stable” asset, often the U.S. dollar. It is used to move dollars disguised as digital tokens to bypass some of the constraints of traditional banking.

IRGC (Revolutionary Guard)

An elite Iranian military and security force with a massive role in the country’s economy. It is heavily sanctioned by the United States and other governments.

Shadow Banking

A network of vehicles, intermediaries, offshore accounts, and financial schemes that are not regulated like traditional banks but are used to move capital, often opaquely or to skirt regulations.

Hawala

An informal money transfer system based on intermediaries and personal trust, widely used in the Middle East and Asia. It allows funds to be moved without passing through the formal banking system.

Mixing / Layering of Transactions

Techniques used to break up and reroute crypto transactions across many different addresses, with the goal of making it harder to identify the source or the final recipient of the funds.

News N°2

Another app “for child safety” that is actually used to spy on anyone, but with a modern twist: malware franchising. KidsProtect is an Android app that poses as a parental control tool, but it is designed to grant complete, invisible access to the victim’s phone, turning every smartphone into a pocket-sized wiretap. The interesting part isn’t just what it does, but how it’s sold: not as simple spyware, but as a turnkey platform for anyone who wants to jump into the private surveillance business with their own brand.

The developers of KidsProtect offer the product on a subscription basis, starting at around $60, with declared support from Android 7 up to the latest versions, including Android 16, ensuring they cover practically every device in circulation. But the real “added value” is the white-label package: buyers can change the name, logo, color, and identity, reselling the malware as if it were their own service, setting prices and marketing as they please. The result: even if one brand is shut down or ends up under investigation, the exact same solution can reappear under a different name the next day, turning law enforcement’s job into an indefinite game of whack-a-mole.

Despite yet another facade “designed to protect minors,” KidsProtect is openly advertised on a hacking forum on the clear web-so no dark web, no mystery: it’s all right there in the open. The ad describes it as built for stability and stealth, even offering a one-day free trial to convince potential “resellers” that the product works. Certain details on the profile of the user promoting it and the app’s screenshots suggest the developer is Greek-speaking, but the geographical origin is almost a colorful footnote compared to the business model. Amidst all this, the article recalls the much more famous case of PhoneSpector/Highster Mobile, shut down by a New York court in 2024, only to show how, for every shutdown, two new platforms pop up doing the same thing under a different name.

From a technical standpoint, KidsProtect does exactly what you’d expect from modern spyware: once installed, it vanishes from the user’s sight and masks itself as a benign system element. Instead of appearing as a parental control app, it shows up under generic names like “WiFi Service” or “WiFiService Installer,” and the accessibility and notification modules take on equally harmless labels, such as “WiFiService Assistant” and “WiFiService Monitor”, perfect for not drawing attention from anyone taking a quick glance at their app list. The package name is com.example.parentguard, a curious choice to say the least, since “com.example” is the prefix usually used in educational examples and templates, not in actual commercial products; it’s a polite way of saying “we don’t want you to know who we are.”

Looking at the permissions requested, it’s hard to distinguish it from a professional surveillance tool: background location access, microphone, camera, SMS, call logs, contacts, and so on. It also abuses the Accessibility Service, used to read any content on the screen and intercept keystrokes, including passwords and various codes, allowing the remote operator to see practically everything happening on the device. Other permissions, such as SYSTEM_ALERT_WINDOW and the request to be excluded from battery optimizations, serve to ensure maximum persistence, preventing the operating system from closing it to save power. Added to this is a BootReceiver that restarts the components every time the phone is rebooted, and a Device Administrator module (MyDeviceAdminReceiver) that makes removing the app through normal Android settings much more complicated.

Even the installation phase doesn’t pretend to be clean: the page distributing the APK clearly instructs the user to disable Google Play Protect, because otherwise, the solution would be flagged as suspicious by Android’s security scanner. In practice, to use this supposed tool for “protecting children,” the first step is to turn off your protection. Analysts who examined the APK have also published the SHA-256 hashes associated with the file, allowing security solutions and network administrators to identify and block the spread of the malware in their environments.

The article closes with some common-sense tips that should be obvious by now, but evidently aren’t obvious enough: always keep Google Play Protect active, avoid installing APKs from unknown sources or fly-by-night websites, and be wary of apps that demand access to the Accessibility Service without a truly clear and explicit reason. It is also worth periodically checking which apps are registered as device administrators in your security settings; if something ambiguous appears with names like “WiFi Service” or you find the package com.example.parentguard, the diagnosis is simple: the phone is likely compromised. In that case, we’re not talking about a “minor issue,” but full-scale surveillance, and the only sensible response is to remove the software immediately, report the incident, and consider a factory reset of the device.

Source: https://cybersecuritynews.com/new-spyware-platform-lets-buyers-rebrand/

Essential Glossary #2

Spyware

Software designed to gather information about a device or a user without their knowledge, often in a continuous and silent manner.

Parental control

A category of applications designed—at least in theory—to allow parents to monitor and limit their children’s device usage (websites visited, apps used, screen time).

White-label

A business model in which a product is sold without a brand, allowing the buyer to resell it under their own brand as if they had created it themselves.

APK

The installation file for Android apps (Android Package); installing APKs from outside the official store often means putting your trust in whoever provided the file.

Google Play Protect

A security system built into Android devices with Google services that scans installed or incoming apps for harmful behavior.

Accessibility Service

An Android feature designed to assist users with disabilities by allowing certain apps to read and interact with screen content; if used maliciously, it can become an open window into everything you do.

Device Administrator

A special privilege that some apps can obtain on Android to gain advanced permissions, such as preventing easy uninstallation or controlling specific security functions.

SHA-256 Hash

A file’s digital cryptographic fingerprint; if two files have the same hash, they are virtually identical. The hash allows for the certain identification of a specific piece of malware.

BootReceiver

An Android app component that runs automatically when the device starts up; it is useful (for spyware developers) to reactivate surveillance software every time the phone is turned back on.

News N°3

North Korea as the Involuntary Sovereign Wealth Fund of Crypto

If anyone was still under the illusion that cryptocurrency was a paradise of financial freedom, the data from TRM Labs offers a more prosaic perspective: three-quarters of everything stolen in the crypto sector during the first four months of 2026 is now resting quietly in Pyongyang’s coffers. Seventy-six percent, to be exact. Not bad for a regime that is theoretically cut off from the global financial system.

The beauty of it is that the North Koreans aren’t even working particularly hard. They aren’t responsible for 76% of the attacks; they’ve simply realized that it’s better to aim for quality over quantity. While other criminal groups waste their breath on thousands of two-bit scams, Kim Jong-Un’s regime lands two well-placed blows and walks away with $577 million in just eighteen days: $285 million from Drift Protocol and $292 million from KelpDAO. That’s a return that would make any Wall Street hedge fund turn pale.

As early as 2017 and 2018, North Korea had sensed something that crypto enthusiasts prefer to ignore: a system designed to escape “oppressive” institutional safeguards is, by definition, a system that cannot protect itself. Banks can block a suspicious transfer to Pyongyang. Blockchains, on the other hand, are often structurally incapable of doing so, and when it would be technically possible, the crypto community actively chooses not to intervene, staying faithful to its anarcho-capitalist dogma even at the cost of its own wallets.

The result is that while the United States thought it was economically strangling the North Korean regime, thousands of amateur traders have built a digital highway perfect for money laundering. In 2025 alone, Americans lost over $11 billion to crypto scams, according to the FBI. The global total is likely much higher, but who really wants to know when you can just believe that the “future of finance” is right around the corner?

According to Ari Redbord of TRM Labs, the recent escalation could be explained by the increasingly massive adoption of Artificial Intelligence by North Korean hackers. AI eliminates the traditional barriers that used to limit the precision of attacks: language difficulties disappear, the time needed to build convincing personas is drastically reduced, and the personalization of attacks becomes scalable. Experts report a 500% increase in AI-assisted scams over the last year.

It’s not just about more sophisticated social engineering. Large language models help synthesize enormous amounts of data, generative tools write code, and the entire process of identifying and exploiting vulnerabilities is compressed from days to hours. Bradley Smith of BeyondTrust hits the nail on the head: “Governance structures that require hours or days will not survive against an AI-powered attacker operating in minutes.”

The structural problem, as Smith explains, is that DeFi platforms manage values equivalent to those of a nation-state using startup security architectures. Single points of trust, no validation of the origin of assets moving between systems, and governance structures incapable of responding at the speed of the attacks. Until the crypto ecosystem imposes the same standards of trust verification that traditional financial infrastructures require, state-sponsored actors will continue to treat it as the lowest-cost financing mechanism available.

The tragicomic thing is that all of this was predictable. In fact, it was inevitable. Building a financial system that rejects every form of supervision on principle and then being surprised that it becomes the favorite ATM for rogue regimes is like leaving your front door wide open and complaining about the burglaries.

Source: https://www.darkreading.com/cybersecurity-analytics/crypto-stolen-2026-north-korea

Essential Glossary #3

APT (Advanced Persistent Threat)

Sophisticated hacking groups, often state-sponsored, that conduct prolonged and targeted attack campaigns against specific high-value objectives.

DeFi (Decentralized Finance)

Blockchain-based financial platforms that operate without traditional intermediaries like banks, utilizing smart contracts to automate transactions and financial services.

Smart Contract

Computer programs that automatically execute agreements when specific conditions are met, eliminating the need for a central authority to verify the execution.

Social Engineering

Psychological manipulation techniques used to trick victims into revealing sensitive information or performing actions that compromise security.

LLM (Large Language Model)

Artificial intelligence models trained on massive amounts of text that can generate natural language, analyze data, and assist in various cognitive tasks.

News N°4

A phishing operation dubbed “AccountDumpling” is exploiting blind trust in major platforms to steal Facebook accounts en masse, specifically business accounts, the ones that are actually worth something. We aren’t talking about your typical sketchy link riddled with typos; we’re talking about emails delivered directly from Google AppSheet servers, digitally signed by the book, and perfectly aligned with the security standards everyone repeats to us like a mantra. Essentially: the message is technically flawless, but the content is a trap.

The trick is as simple as it is effective: attackers use legitimate AppSheet notifications to send communications from addresses like “noreply@appsheet.com,” which breeze past corporate filters, security checks, and red flags. The targets are Facebook page administrators, business accounts, and profiles linked to ad campaigns and credit cards, the kind of accounts that, if compromised, provide access to cold hard cash, real audiences, and serious reputational damage. The messages claim policy violations, imminent suspensions, blue badge verifications, or urgent account issues. Sometimes they even dangle job or collaboration opportunities, because as we all know: when a professional ego is involved, the guard rails come down.

Once the victim takes the bait, they are dumped into a diverse ecosystem of phishing pages: Netlify sites mimicking Meta help centers, Vercel panels promising blue badges or special verification, and Google Drive PDFs professionally designed to look like official correspondence. Some even use fake recruiters who move the conversation to more “human” channels like WhatsApp or external chats. At every step, the request remains the same, though packaged with increasing gravity: credentials, 2FA codes, personal data, ID documents, and phone numbers, all under the usual guise of “security verification” or “account recovery.” The whole thing is seasoned with layers of code obfuscation and anti-analysis techniques, because even criminals don’t want to be studied too closely.

The harvested data doesn’t just sit there gathering dust; it flows in real-time to Telegram bots and channels. The operation managers receive each new victim like a notification for a new order on an e-commerce site. Nicknames like “Big Bosss” and bots with explicit names, often in Vietnamese, point to an organized structure that treats compromised accounts as ready-to-use or resaleable merchandise. According to data analyzed by researchers, we are looking at at least 30,000 compromised accounts, with data streams still flowing, proving that this campaign is no isolated experiment but a full-scale infrastructure in peak operation. Most victims are in the United States, but the list of affected countries is long and includes Italy, confirming that no digital market is truly “marginal” when there’s money to be made.

When it comes to attribution, the signs are fairly clear: document metadata featuring Vietnamese names, language-specific code comments, activity hours consistent with local time zones, and linguistic references in bot names like “Tích Xanh,” which directly refers to the “blue checkmark.” There’s even a reference to a certain “Satiated Cat” in the code of some variants, likely the handle of the developer who creates the kits used by others for the operational side. The result is an industrial model: someone writes and sells the kits, others use them to steal accounts, and still others exploit or resell those accounts in shadow markets for fake ad campaigns, scams, and disinformation.

The ultimate paradox is that all of this works precisely because the security ecosystem has spent years training users to trust certain details: emails with valid SPF, DKIM, and DMARC records, reputable domains, hosting on “respectable” services, polished graphics, and Google Drive PDFs. Here, those very signals are flipped and weaponized: it’s no longer the sketchy link that gives the criminal away, but the “too clean” context that disarms all suspicion. Some victims contacted by researchers spoke of sudden account lockouts, unauthorized activity, maxed-out credit cards, and ad campaigns launched without their knowledge, resulting in financial losses that are anything but symbolic. In other words, this isn’t just the usual story of “don’t click on strange links”: here, the problem is that the link doesn’t look strange at all.

Source: https://guard.io/labs/accountdumpling---hunting-down-the-google-sent-phishing-wave-compromising-30-000-facebook-accounts

Essential Glossary #4

Phishing

An attack technique where someone poses as a legitimate service (bank, social media, platform) to trick you into voluntarily handing over credentials, codes, and personal data.

Business Account / Facebook Business Page

Profiles and pages used for commercial activities, marketing, and advertising, often linked to credit cards and ad budgets. Stealing these is far more profitable than stealing a random private profile.

Google AppSheet

A Google service that allows users to create apps and workflows without writing code; here, it is exploited to send “clean,” legitimately signed emails used as a Trojan horse.

SPF, DKIM, DMARC

Technical standards used to verify that an email actually comes from the domain it claims to use. In an ideal world, they help block spam and scams; in this case, they make malicious emails look more credible.

Netlify / Vercel

Legitimate hosting and deployment platforms for websites and applications. Attackers use them to host fake copies of Meta/Facebook pages to look even more “professional.”

2FA (Two-Factor Authentication)

A second layer of security beyond the password, usually based on a temporary code. If you hand it over to a phisher along with your password, you’re essentially giving them the spare key.

Code Obfuscation

A technique used to make code difficult to read and analyze, even though it remains technically functional. It is used to complicate the work of investigators trying to figure out what a page actually does.

Telegram Bots / Channels

Tools within the Telegram messaging app that allow for the automated collection and distribution of data. In this context, they act as an operations center to receive stolen credentials and info in real-time.

Blue Badge / “Tích Xanh”

The verified account symbol on social media platforms, used as psychological bait to lure victims into “verification” procedures that are actually designed to steal their accounts.

Phishing Kit

Ready-to-use packages that include clone pages, scripts, Telegram integrations, and instructions. They allow even low-tech criminals to run complex campaigns by leaning on the work of the original developers.

News N°5

Yet another story of how a “state-sponsored” espionage operation manages to infiltrate half a government, only to go belly up because someone forgot an open directory on a two-bit VPS in the Emirates. An operation linked to an Iranian-nexus actor worked quietly on at least 12 Omani ministries, walking away with tens of thousands of records and leaving behind convenient persistent access for whenever they feel like dropping by the systems again.

Everything surfaced when the staging server 172.86.76[.]127 was found with its entire arsenal exposed: toolkits, command-and-control code, session logs, stolen data, and even a README labeling the machine as “VPS C2”, just so there was no doubt about its function. The top-tier target is the Ministry of Justice and Legal Affairs (MJLA), which remained compromised at least until April 10, 2026. Meanwhile, entities like the Royal Oman Police, the Tax Authority, the Civil Aviation Authority, the Ministry of Finance, and the Public Prosecutor’s Office also appear to have been hit. The analysts at Hunt.io, having found the “backroom door” wide open, simply documented the disaster: tools, targets, and the volume of outgoing data, all following a textbook regularity.

The technical core of the operation is depressing in its simplicity: two ASPX webshells, hc2.aspx and health_check_t.aspx, which accept commands via HTTP parameters, pass them to the system through Windows processes, and spit out plaintext output back to the attacker. On the server, there is also a folder containing 12 exploit scripts dedicated to Omani government targets, covering credential spraying on Exchange, SQL Server escalation, and in-memory execution techniques to avoid leaving files on the disk. To gain entry and move laterally, they used the “evergreens”: ProxyShell, credential brute-forcing, and GodPotato to gain elevated privileges once inside, because if something keeps working, there’s no need to invent anything brilliant.

Over 26,000 user records were extracted from the Ministry of Justice systems, along with information on court cases, committee decisions, and Windows registry hives containing internal credentials, exactly the kind of material no government wants circulating outside its borders. The command-and-control component runs on a Python HTTP server, with a PowerShell beacon on the victim’s side checking in every 30 seconds, transmitting domain, user, hostname, and stolen data in small encoded chunks to avoid detection, at least in theory. Logs show active sessions on April 10, 2026, starting at 03:00 UTC, all traced directly back to the Ministry of Justice network, while over 26,000 MJLA records simultaneously appear in the exfiltration logs.

The infrastructure is not isolated: within the same hosting network, domains appear that mimic Persian-language media and services related to censorship-circumvention tools, a pattern already seen in other operations attributed to the Iranian state sphere. The tooling shows overlaps with Iranian-nexus groups like APT34 and MuddyWater, which have long targeted Middle Eastern governments using similar techniques, though Hunt.io limits itself to a “regional” attribution without pinning a specific label on a single group. On the infrastructure front, several IPs in the 172.86.76[.]x range pop up (”dubai-*.vaermb[.]com” series), plus IP 45.59.114[.]60 linked to exceptionnotfound[.]ir domains, all sitting behind providers like RouterHosting and, in some cases, Cloudflare. The IoCs are appropriately “defanged” with [.], with a polite recommendation to “refang” them only in controlled environments like MISP, VirusTotal, or corporate SIEMs to prevent anyone from clicking them from a production PC.

In the end, the moral is cliché but timely: monitoring exposed infrastructure, in the brief window before attackers clean up their tracks, remains one of the most practical ways to realize your systems are being stripped down while you think you have everything under control. In Oman, this monitoring arrived late, and the result is a mix of webshells, SQL escalation, and data theft that perfectly illustrates how little it takes today to turn ministries and authorities into a data open bar for those with time and patience.

Source: https://cybersecuritynews.com/iranian-nexus-operation-targets-oman-ministries-with-webshells/

Essential Glossary #5

Iranian-nexus threat actor

A label used to identify attack groups linked, directly or indirectly, to the Iranian state apparatus, without having to commit to a specific group name.

Webshell

A small script (often in ASPX, PHP, etc.) installed on a compromised server that allows the attacker to execute remote commands via a browser, as if they had a shell open on the system.

SQL Server escalation / Privilege escalation

Techniques used to move from limited access to one with higher permissions (administrator, system) by exploiting vulnerabilities or weak configurations in the SQL database or operating system.

ProxyShell

A chain of known vulnerabilities in Microsoft Exchange that allows for remote code execution on the mail server; it isn’t new, but it continues to work wherever patches haven’t been applied.

Credential brute-forcing / Password spraying

Automated attempts to guess passwords by trying many combinations or using the same “weak” password across many different accounts until something gives.

GodPotato

A Windows exploit tool that leverages system service weaknesses to obtain elevated privileges, often reaching the equivalent of “administrator” or “SYSTEM.”

C2 (Command and Control)

Infrastructure used by attackers to communicate with compromised machines: they send commands, receive output and stolen data, and coordinate the operation remotely.

Beacon

A component (for example, a PowerShell script) installed on the victim’s machine that “checks in” at regular intervals with the C2 server, announcing that it is active and ready to receive instructions.

IoC (Indicator of Compromise)

Technical forensic evidence of an intrusion: IPs, domains, file hashes, webshell names, etc., used to identify or block an attack within logs and security systems.

Defang / Refang

A technique to make an indicator harmless (e.g., writing example[.]com instead of example.com) to prevent accidental clicks; “refanging” means restoring it to its normal format for lab analysis.

News N°6

Daemon Tools, that “harmless” little program for mounting disc images that many installed without a second thought, has become yet another Trojan horse distributed directly from the manufacturer’s own stable. Since roughly April 8, official servers from developer AVB have been serving perfectly signed installers, appearing entirely legitimate, that infected the Daemon Tools executables and planted a nasty piece of malware that launches at Windows startup (versions 12.5.0.2421 through 12.5.0.2434).

The result is that thousands of machines across over 100 countries have ended up under silent surveillance, while users thought they were doing the right thing by updating their software from the official site. The first malicious component doesn’t do anything flashy: it sticks to gathering system info, MAC addresses, computer names, DNS domains, active processes, installed software, and local settings, sending it all to a server controlled by the attackers—in a perfect “I’m watching you, then I’ll decide if you’re worth my time” style.

The real “fun” starts later: out of all the infected machines, only about a dozen in the retail, scientific, government, and manufacturing sectors received an extra payload, a clear sign that this isn’t just some run-of-the-mill, sloppy ransomware, but a targeted operation. One of the additional payloads is a minimalist backdoor, the digital equivalent of a foot in the door: it executes commands, downloads files, and runs shellcode directly in memory to leave as little as possible on the disk. In at least one case involving a Russian educational institution, something more sophisticated was spotted: a backdoor called QUIC RAT, capable of injecting code into legitimate processes like notepad.exe and conhost.exe and communicating with its command-and-control server using practically any protocol it can get its hands on (HTTP, UDP, TCP, WSS, QUIC, DNS, HTTP/3).

Geographically, most of the hit organizations are concentrated in countries like Russia, Brazil, Turkey, Spain, Germany, France, Italy, and China. Kaspersky, which discovered the mess, admits its visibility is limited to systems where its products are installed, so the real picture could be even worse. According to their analysis, about 10% of the machines belong to companies and organizations, and the more complex backdoor was distributed only to a small group of “interesting” targets, with motives that are still unclear: espionage, “big game hunting,” or both.

Of course, this isn’t an isolated incident, but just the latest episode in a series that would be quite boring if it weren’t so devastating: CCleaner in 2017, SolarWinds in 2020, 3CX in 2023, followed by supply chain attacks on tools like Trivy, Checkmarx, Bitwarden, and over 150 open-source packages. All cases carry the same unsettling moral: you can be as meticulous as you want, but if the “official” channel is compromised, installing a signed update becomes practically an act of faith. And like clockwork, it takes weeks or months before anyone notices the distribution chain has been poisoned.

The final recommendations are what we’d expect once the damage is already done: anyone using Daemon Tools should scan their entire system with a serious antivirus and, on Windows, check the indicators of compromise published by Kaspersky. More tech-savvy users should keep an eye out for suspicious code injection in legitimate system processes, especially those starting from executables launched in “open” folders like Temp, AppData, or Public. In essence, it’s the usual call for forensic investigations on machines that until yesterday were considered “clean” simply because the software came from the right site and carried the right signature.

Source: https://arstechnica.com/security/2026/05/widely-used-daemon-tools-disk-app-backdoored-in-monthlong-supply-chain-attack/

Essential Glossary #6

Supply-chain attack

An attack where hackers don’t target you directly, but rather one of the links upstream (a software provider, library, or service) that you use and trust. This way, the malware reaches you disguised as an “official update.”

Backdoor

A hidden mechanism that allows whoever installed it to re-enter the machine at will, execute commands, and download or upload files—often without leaving a trace visible to the user.

Payload

The “cargo” or content of the attack; it is what the malware actually does once inside. It can be a simple info-stealer, a full-blown backdoor, ransomware, and so on.

Shellcode

A small block of code executed in memory, often used to launch or control other parts of an attack while minimizing the number of visible files left on the disk.

C2 (Command and Control)

The infrastructure (servers and protocols) that malware uses to communicate, receive commands, and send stolen data. It is the attackers’ “mission control.”

QUIC RAT

An advanced backdoor observed in this attack, capable of injecting itself into legitimate processes and using various protocols (including QUIC and HTTP/3) to talk to the C2 server, making it harder to block or track.

MAC address

A unique identifier associated with a device’s network interface. It is often used to precisely distinguish one specific machine from others.

Indicators of Compromise (IoC)

Technical footprints that suggest a system has been breached: suspicious files, hashes, domains, IP addresses, anomalous log patterns, and so forth.

Digital signature (Installer)

A mechanism where the manufacturer “signs” their software to prove it hasn’t been tampered with. However, when attackers compromise the signing process or the distribution server directly, this guarantee becomes unreliable.

However, there is a way to keep your hands glued to the keys, letting the page follow your lead as you type. Try this little experiment: start a new line, type a simple asterisk (*) and immediately press the spacebar. Without you having to do anything else, that small symbol transforms and shifts slightly, becoming a perfect, tidy bullet point.

It’s an automation that brings a strange sense of satisfaction, as if the blank page understood exactly where you were going before you even finished the word. The same thing happens if you prefer to list facts by importance or chronology: type the number one, add a period, press space, and Google Docs realizes you’ve started a sequence, taking care of the formatting for you.

Instead of wrestling with drop-down menus or hunting for icons that always seem to hide when you need them, you can let your post’s structure organize itself. This little shortcut acts like a silent assistant cleaning your desk while you work on it, allowing you to focus solely on the clarity of what you want to say, without ever taking your eyes off the heart of your story. It’s a much more agile way to bring order to the chaos of your thoughts, turning the keyboard into a tool that supports your natural movement instead of holding it back.

Subscribe now

Area 1: Setup and Basecamp (Initial Phase)

Optimize AI behavior to save time and oxygen.

Area 2: Analysis and Senses (The Ascent)

Use multimodality to see through the data fog.

Area 3: Engineering and Automation (The Summit)

Commands for true software architects and autonomous agents.

PRO TIP: The Magic Delimiter

When mixing instructions and complex data, always use XML tags to avoid confusing the agent.

<context> Paste your book or code here </context>

<instruction> Perform a risk analysis based on the context above </instruction>

⚙️ Area 4: Maintenance and Fine-Tuning (High Altitude)

Manage computing power like a true veteran.

• For massive queries: “Set thinking_level to High and analyze the entire repository for vulnerable dependencies.”

• To save budget: “Cache this document context for the next 4 hours (TTL 14400s) and respond to subsequent queries in Flash mode.”

Technical Glossary: The Words of the Summit

A

Agent Manager: The central module of Google Antigravity that coordinates various autonomous agents, assigning tasks such as writing code, testing, or documentary research.

Hallucination: A logical error where AI generates plausible but false information. It is countered through Grounding.

Antigravity: Google’s agentic development environment (IDE) that allows software to be built through the coordination of AI agents instead of manual line-by-line writing.

API Key: A unique digital key that allows external applications (like N8N) to “talk” to Gemini, acting as a pass for its services.

Artifact: A dynamic document generated by agents (e.g., a Task List or an implementation plan) that documents work progress within Antigravity.

C

Canvas: An interactive workspace alongside the chat, ideal for writing long texts or code, allowing for surgical edits and format transformations.

Context Caching: A technique that allows for temporarily storing large amounts of data (up to millions of tokens) to reduce costs and latency in subsequent queries.

Context Window: The maximum amount of information Gemini can “keep in mind” at one time. With Gemini 3, this window reaches one million tokens.

D

Deep Research: An advanced research tool that navigates dozens of web sources to create exhaustive reports, comparative tables, and verified bibliographies.

Delimiters (XML Tags): Symbols like <instructions> or </text> used to clearly separate commands from data within a complex prompt.

F

Few-shot Prompting: An instruction technique that involves providing the AI with a few concrete examples (shots) of the desired result to guide its style and precision.

G

Gem: A customized AI assistant trained on specific instructions and documents to perform a vertical role (e.g., Language Tutor or SEO Analyst).

Gemini Flash: The version optimized for speed and lightness, ideal for quick tasks and massive automations.

Gemini Pro / Advanced: The high-power version, optimized for complex logical reasoning, mathematics, and programming.

Grounding: The process that forces AI to base its answers exclusively on certain sources (uploaded documents or Google Search), reducing hallucinations.

M

MCP (Model Context Protocol): A standardized protocol that acts as a “bridge” to connect AI to local databases, system files, or external APIs in real-time.

Multimodality: Gemini’s native ability to understand and process text, images, audio, video, and code simultaneously.

N

N8N: An automation platform that allows connecting Gemini to hundreds of other apps to create autonomous workflows.

Nano Banana Pro: The model specialized in advanced image generation and editing (including the Face Swap function).

R

RAG (Retrieval-Augmented Generation): A system that enhances AI responses by allowing it to consult an external knowledge base (like a NotebookLM) before responding.

S

Skill: A modular extension for Antigravity agents containing instructions and scripts to execute specific tasks (e.g., web publishing).

T

Thinking Level: A parameter that regulates the AI’s cognitive effort, allowing a choice between quick responses or deep, multi-step reasoning.

Thought Signature: A trace of the logical path followed by the AI during reasoning, useful for verifying the coherence of decisions made by the model.

Token: The basic unit of measurement for information processed by AI (about 4 characters of text). It is the “fuel” that determines the cost and memory capacity of the model.

V

Veo 3: Gemini’s cinematic model capable of generating high-definition videos complete with audio and environmental sound effects.

Vibe Coding: An approach to programming based on describing the idea and interface in natural language, leaving the task of generating the entire software to the AI.

Conclusion: The Summit – Looking Toward the Horizon

You’ve made it. You’ve planted your flag on the highest peak of the Gemini ecosystem. Looking down, you can see the entire trail: from those first uncertain steps at Basecamp, where you learned to adjust your pack with custom instructions, through the vertical walls of the Ascent, where you activated your multimodal senses and created your first Sherpas (Gems). Finally, you mastered the Summit, building digital bridges with the MCP and delegating the heavy lifting to autonomous agents in Antigravity.

The View from the Top: AI as an Extension of the Intellect

Reaching the summit doesn’t mean the challenge is over. The true insight of an expert climber is understanding that artificial intelligence is not a substitute for human thought, but an enhancement of it.

• The Human at the Center: Even with the most advanced N8N automations or the lightning-fast reflexes of Gemini Flash, the direction, the ethics, and the “why” of the climb remain in your hands.

• Continuous Evolution: The AI mountain is volcanic; it changes shape every day. New models, more refined parameters, and deeper integrations will emerge, but the foundations of Prompt Design and Agentic Logic you’ve acquired are your universal compass.

The Climber’s Logbook

Here are the three principles to carry with you as you begin the descent toward new challenges:

1. Experiment Without Fear: Use Vibe Coding to bring your wildest ideas to life. A prototype failing in AI Studio is just a handhold that gave way, not the end of the expedition.

2. Stay Anchored to the Truth: Remember the importance of Grounding and NotebookLM. In a world of infinite data, accuracy and citing sources are your safety rope.

3. Automate Value, Not Thought: Delegate repetitive tasks to free up mental space. The true summit is having the time to imagine the next mountain to climb.

PRO TIP: Never climb alone. The community of AI explorers is your greatest resource. Share your Artifacts, publish your N8N workflows, and keep exchanging climbing techniques. Knowledge, when shared, makes the mountain less steep for everyone.

VISUAL ARCADE GUIDE

Chapter “Cheat Sheet” Concept: A summary poster featuring magic “arcade-style” commands.

Image Generation Prompt: Pixel Art, 8-bit style, 1980s “Cheat Code” screen. A black background with a phosphor green grid. In the center, a pixelated sheet of paper with glowing text representing the most powerful prompts. Around the sheet, glowing power-up icons (lightning bolts, stars, diamonds) shine. At the top, the title “ULTIMATE CHEAT SHEET” in a yellow arcade font. Clean, readable, and iconic style.

Chapter “Technical Glossary” Concept: An illustrated index of all treasures and tools collected.

Image Generation Prompt: Pixel Art, 16-bit arcade style. An “Inventory” screen from an 80s RPG. In the center, a grid of glowing icons: a scroll (Prompt), a blue potion (Token), a gold key (API Key), a lightsaber (Gemini Pro), and a light bridge (MCP). At the top, the words “TECH GLOSSARY” in a neon pixelated font. In the background, the silhouette of the climbed mountain under a starry sky. Vibrant colors: gold, cyan, and purple.

Conclusion Concept: Achieving the final goal, the celebration, and the new horizon.

Image Generation Prompt: Pixel Art, 16-bit “Ending Screen” aesthetic. The pixelated protagonist stands on the highest peak, silhouetted against a golden and purple sunset. They hold the Gemini flag planted in the snow. In the foreground, the game credits roll slowly upward. On the horizon, the silhouettes of other even higher and more technologically advanced mountains can be seen. At the bottom, the flashing text “CONGRATULATIONS - A NEW ADVENTURE AWAITS”. Warm and vibrant colors, triumphant arcade music visually suggested by pixelated musical notes in the air.

News n°1

The company that was supposed to protect civil liberties is now helping to... do the exact opposite. Welcome to the Palantir paradox, where employees are starting to wonder if they’ve accidentally ended up on the wrong side of history.

The betrayal of the original mission

Palantir was born with CIA funding after 9/11, with a goal that sounded almost noble: defending civil liberties. Eighteen years later, with Trump back in the White House, the company has become the DHS’s technological infrastructure for deporting immigrants. A former employee summarized the 180-degree turn with elegance: the threat “is now coming from within,” and the company that was supposed to prevent abuses “now seems to enable them.” In short, they went from defenders to executioners, just with better paychecks.

When the company chat gets awkward

In January 2026, a nurse named Alex Pretti was killed during protests against ICE. Palantir employees started asking uncomfortable questions. Management’s response? Automatically deleting Slack conversations after seven days in the #palantir-in-the-news channel. The official reason cited “information leaks,” but translated, it means: it’s better not to leave a paper trail of collective unease.

Things got worse on February 28, when a missile strike on an Iranian elementary school killed over 120 children, and it turned out that surveillance systems like Palantir’s Maven were used. At that point, even Slack’s auto-cleanup wasn’t enough to contain the discontent.

The manifesto of cognitive dissonance

During an AMA session in February, a member of the privacy team candidly admitted that “a sufficiently malicious client is practically impossible to prevent right now” and that attempts to talk sense into CEO Alex Karp have been “largely unsuccessful.”

Karp, meanwhile, published a manifesto based on his book in which he proposes reintroducing the draft. Employees complained because such statements make it harder to sell the software outside the U.S. Karp’s response was a masterpiece of CEO rhetoric: “if you have a position that never costs you a single employee, it’s not a position.” In other words: if you don’t like it, there’s the door.

It’s hard to find a more “elegant” way to tell your employees that their ethical concerns are worth less than the corporate vision. And so, the company that was supposed to defend freedom continues undeterred to do the opposite, while those working there discover that maybe, after all, they really are the villains of the story.

Source: https://attackofthefanboy.com/tech/palantir-dropped-its-infamous-manifesto-a-while-back-and-now-its-employees-are-starting-to-question-whether-theyre-the-bad-guys/

News n°2

Meta and Microsoft have just given a crystal-clear lesson on how AI capitalism works in 2026: profits are soaring, Wall Street loves the stock, and those heading out the door are simply the “transition cost” toward a future where an Excel sheet carries more weight than any “talent.”

Meta is laying off 8,000 people (roughly 10% of its workforce) and canceling another 6,000 open positions, all effectively cut as of May 20.

Microsoft, on the other hand, avoids using the word “layoff”: it is offering the first “voluntary retirement” program in its history to up to 8,750 U.S. Employees, about 7% of its staff, provided that their age + years of service equal at least 70.

The result is identical: up to 23,000 roles are either vanishing or will never be filled, even though both companies are posting record-breaking revenue and profits. Economic hardship has nothing to do with it; quite simply, they’d rather spend their money elsewhere.

The real budget: machines, not people

Meta expects to burn between $115 and $135 billion in capex in 2026, nearly double the $72 billion spent in 2025, virtually all of it destined for data centers, Nvidia GPUs, custom chips, infrastructure for Llama, and the new Meta Superintelligence Labs division.

Meanwhile, in the fourth quarter of 2025 alone, it posted a net profit of $22.8 billion on $201 billion in annual revenue: no sign of a company “struggling.”

Microsoft follows the same logic, just a more polite version: quarterly revenue at $81.3 billion (up 17% year-over-year), Azure at +33%, with AI services accounting for 16 points of that growth, and hundreds of billions promised to data centers, Copilot, and the OpenAI partnership.

The generous exit packages for senior staff become a mere statistical detail in an AI spending plan that turns cut salaries into background noise.

Meta is proceeding with blunt layoffs across all business units while shifting engineers toward a new Applied AI division and toward its small-business advertising arm.

The corporate narrative has slid from “we have too many low performers” (2025) to “matters of efficiency and contribution,” which implicitly means: we’re not kicking you out because you’re bad at your job, but because you’re in the wrong department for the story we need to tell investors.

Microsoft, more refined, hides behind the “Rule of 70,” which effectively targets those in their 50s and 60s who built the pre-AI Microsoft.

The hardest-hit departments are precisely those where automation via Copilot and AI agents is most advanced (Cloud, Gaming, Global Sales), following a 2025 spent ramping up performance pressure and multiplying PIPs.

Outside of Meta and Microsoft, the pattern is the same: Oracle has cut up to 30,000 roles (about 18% of its workforce) to shift $8-10 billion a year toward a $156 billion AI infrastructure plan, while its remaining contractual obligations reach $523 billion and profits have nearly doubled.

Amazon has restructured 16,000 positions, Dell 11,000, and Snap has cut 16% of its staff: in 2026, over 96,000 tech workers have already lost their jobs, 40% more than in the same period in 2025.

Since 2020, tech layoffs have neared 900,000 positions, with a narrative that has evolved with suspicious convenience: first “we over-hired during Covid,” then “we are restructuring for AI,” and now “we are cutting for AI”, stated explicitly, as if it were an act of modernization.

It has become a sort of magic formula: say “AI” and everything, from staff cuts to collapsing internal morale, suddenly becomes rational and inevitable.

The trendy term is “AI-washing”: layoffs announced in the same breath as a solemn nod to AI strategy, without any proof that internal systems are actually capable of doing the work of the people being sent home.

New divisions like Superintelligence Labs and Meta Compute exist in press releases far more than in operational processes, but they are enough to create the illusion that Meta’s $115-135 billion in capex or Microsoft’s $81 billion in annual investments are “naturally” replacing thousands of salaries.

The bottom line: it’s not the technology, it’s the hierarchy of values

From 2022 to 2026, we’ve seen three waves: first the post-pandemic excuse, then restructuring “for AI,” and now the direct, unapologetic link between cuts and AI investment.

Oracle openly says it lays people off to build data centers, Meta says it cuts to “offset” AI investments, and Microsoft designs a program tailored for those who are no longer aligned with its AI-centric future.

The slaughter has only just begun.

Source: https://thenextweb.com/news/meta-microsoft-layoffs-23000-ai-spending

News n°3

Putin has found a new way to get under Russians’ skin: intermittently shutting down the internet, especially mobile data, under the guise of security and the war effort.

The result: people can’t work, communicate, or access basic services, and they are suddenly realizing that the promised “stability” comes with a very concrete price tag in their daily lives.

It’s no longer just dissidents and activists complaining; now it’s influencers, starlets, and typically apolitical figures who, until yesterday, were hawking face creams and a patriotic lifestyle, but today are openly saying that life in Russia has become unlivable.

One of their posts goes viral, forcing the Kremlin to respond and promise vague “course corrections,” as if the whole thing were just a minor technical glitch.

Meanwhile, the state is pushing citizens toward the national super-app MAX, controllable and unencrypted, which is a great idea if your goal is to turn everyone’s phone into an extension of the censorship office. VPNs, workarounds, hacks: many Russians are trying to fight back, but between blackouts and throttling, even digital resistance has its limits.

Official polls, the ones that usually show undying love for the leader, are starting to dip: Putin’s approval rating is sliding back toward pre-invasion levels, while irritation and fatigue over the war, taxes, and online blocks continue to grow. Messing with the internet, essentially hitting people where they live, seems to hurt his approval ratings more than a thousand patriotic TV appeals ever could.

Even the “controlled opposition” is waking up: Communists and new pro-government parties are openly criticizing the blackouts, feigning surprise at a system they themselves have legitimized for years.

Some venture comparisons to Tsarist Russia, speaking of growing discontent and the risk of “more decisive steps,” but in practice, no one is truly challenging the top tier of power.

Underneath it all remains a country where more and more people view elections as a charade, speak of Russia as an open-air prison, and no longer believe the Kremlin’s promises to fix things.

For now, the regime isn’t collapsing, but it has discovered an uncomfortable truth: you can control the media, but once you start cutting the connection, even the faithful begin to wonder if you’ve finally gone too far.

Source: https://www.nytimes.com/2026/04/28/world/europe/russia-internet-restrictions-putin.html

News n°4

This isn’t some epic struggle over the fate of humanity; it’s just a run-of-the-mill settling of scores between two ultra-wealthy men accusing each other of being greedy, cynical traitors to the cause of “the greater good.” Honestly, it’s better than anything on Netflix.

OpenAI was founded in 2015 as a “different kind of lab,” a non-profit project designed to keep Artificial Intelligence from falling entirely into the hands of predatory corporations, driven by the rhetoric of “we’re doing this for humanity, not for profit.” Elon Musk provides the seed money, Sam Altman plays the charismatic leader, and on paper, the idea is noble: a sort of Manhattan Project for AI, but without the part where anyone makes an obscene amount of money. Too bad that right from the jump, Altman writes to Musk suggesting they bump researcher salaries by $100,000 to $200,000 “to signal that we’ll take care of them.” In other words: sure, it’s a non-profit, but with luxury startup pay.

From there, the writing was on the wall. The “non-profit” status lasted about as long as a press release. OpenAI transitioned to a for-profit structure, controlled by employees and institutional investors, Microsoft, Amazon, Nvidia, SoftBank, and a linked foundation, and is now gearing up for one of the biggest IPOs in history, with a valuation nearing a trillion dollars. Altman keeps repeating that he has no direct equity in OpenAI, yet he remains comfortably a billionaire thanks to other investments. Meanwhile Musk, the man portraying himself as the betrayed altruist, is doing exactly what he claims to loathe: building his own AI venture, xAI, within a highly profit-oriented company like SpaceX, which is also destined to go public.

The lawsuit is, officially, a matter of principle. Musk accuses Altman and President Greg Brockman of “unjustly enriching themselves” by billions while betraying OpenAI’s original charter, which spoke of benefiting humanity rather than lining the pockets of a few individuals and major shareholders. OpenAI fires back like they’re in a high-end HOA dispute: you’re not the good guys, he’s the greedy one. The company claims Musk abandoned the project only after realizing he wouldn’t be able to seize total control. Both sides compare themselves, with admirable modesty, to a Shakespearean drama: Musk speaks of “perfidy and deceit of Shakespearean proportions,” while Altman notes that the AI sector is now full of “Elizabethan tragedy-level drama.” The explicit reference is to Julius Caesar: the friend who stabs you in the back, power that corrupts, except here, the blood is replaced by emails and stock options.

The backdrop for all this is the 2026 version of the Regulatory Wild West. Congress is essentially paralyzed, federal watchdogs are weakened, and the Trump administration is packed with people very receptive to the money and glamour of Silicon Valley, the President included. In the absence of effective laws, civil courts are all that’s left: lawsuits become the only way to squeeze out some information and put even a semblance of a boundary on tech power.

Tech regulation activists view the situation with mixed feelings. On one hand, many want OpenAI to be held accountable for its choices just as much as Meta or YouTube. On the other, there’s very little faith in turning billionaires’ personal vendettas into the primary engine for “legislation via private litigation.”

The irony is that, depending on who wins, the outcome is equally bleak. If Musk prevails, OpenAI could be severely weakened or even dismantled, opening up a massive market share for one of Musk’s other companies to “swallow up” without much trouble. If OpenAI manages to get the case dismissed, the takeaway would be just as educational: it’s perfectly fine to start as a reassuring non-profit, promise to work for the public good, and then cynically flip into a trillion-dollar for-profit corporation with zero serious consequences.

Outside the courtroom, the tech world keeps racing: OpenAI is scaling projects ahead of its massive IPO, Google is pledging tens of billions to Anthropic, chipmakers are seeing double-digit growth, and AI is being shoved into everything from retail stores to prison chatbots. Inside the courtroom, however, old 19th-century statutes and apocalyptic fears about the future of AI clash in a trial that won’t make anyone a better person, but might at least force us to look at the structural hypocrisy of an industry that pretends to save the world while filling out an Excel sheet of its profits.

Source: https://www.nytimes.com/2026/04/28/technology/elon-musk-sam-altman-trial.html

News n°5

Yet another miraculous “document reader” has popped up on Google Play, one that did very little reading of documents but could certainly read plenty of cash right out of your bank accounts. Hidden behind that harmless icon was Anatsa, one of the most persistent Android banking Trojans of recent years. The app, published under the package name com.groundstation.informationcontrol.filestation_browsefiles_readdocs, easily blew past the 10,000-installation mark before anyone at Google decided to wake up and pull the plug.

Anatsa is hardly a new kid on the block; it’s been lurking since 2020, specializing in stealing credentials, logging keystrokes, and orchestrating fraudulent transactions while you think you’re just checking your balance. In the meantime, it has expanded its “client portfolio,” managing to target over 831 financial institutions worldwide, including banks and crypto platforms in countries like Germany and South Korea. The latest round was unmasked by the Zscaler ThreatLabz team, who found the app on the Play Store, analyzed it, and published their report in late April 2026, by which time, for the users, the damage was already done.

The strategy is simple, and it works precisely because it leans on a convenient assumption: “if it’s on Google Play, it must be safe.” The app arrives on the store seemingly clean: no embedded malicious payload, no suspicious behavior during the review process. Once installed, it acts like a normal file/document reader just long enough to earn your trust, while behind the scenes, it contacts a remote server and downloads the real cargo: the Anatsa Trojan, installed silently, without any theatrical pop-ups. It’s all thanks to a two-phase delivery designed specifically to bypass static analysis performed only on the app’s initial version.

Once the payload kicks into gear, its first move is to ask for accessibility permissions, masquerading them as necessary for the app to function properly. With that authorization, the malware gains a disproportionate level of control: it can overlay content on other apps, read SMS messages, show full-screen windows, and generally insert itself into any “sensitive” interaction without being noticed. This is how they get to your banking credentials: the user opens their banking app, and Anatsa places a fake login form on top, visually identical to the original, collects the username and password, and ships them off to command-and-control servers.

To top it off, Anatsa also checks its surroundings: if it detects an emulation or sandbox environment, it sticks to showing a perfectly respectable file manager interface, never revealing its criminal side. The result is that automated analysis systems see a harmless app, while on real smartphones, the Trojan goes about its business undisturbed.

In hindsight, the advice is the same stuff that’s been repeated for years but continues to go ignored until something goes wrong. An app pitching itself as a document reader has zero credible reason to ask for accessibility or SMS access; if it does, the correct response is “no,” followed by an immediate uninstall. Keeping Google Play Protect active, avoiding unknown developers, and being wary of “magic” utilities with five stars and two lines of generic reviews is the bare minimum. Anyone who installed the offending app should remove it immediately and scan their device with a decent security tool, not the usual ad-riddled “cleaner” junkware.

Source: https://cybersecuritynews.com/fake-document-reader-on-google-play/

News n°6

Ah, browser extensions. Those incredibly convenient apps that promise to block ads, enhance Netflix, or automate your job search. Too bad that while you think you’re getting a free service, they’re making the deal of the century by selling every single detail of your digital life to anyone willing to pay. And yes, it’s all perfectly legal.

LayerX Security did the dirty work of analyzing the privacy policies of thousands of extensions, discovering that at least 82 of them, with over 6.5 million users, are happily turning your data into cold, hard cash. But here’s the kicker: that’s only what they could actually verify, because 71% of the extensions in the Chrome Web Store don’t even bother to publish a privacy policy. Translation: they only analyzed 29% of the total, which means the number of extensions selling data could be in the tens of thousands.

Take the QVI empire, a network of 24 extensions for streaming platforms like Netflix, Disney+, and HBO Max, published by HideApp LLC. With 800,000 unsuspecting users, they collect viewing history, content preferences, age, and gender, only to resell the full package to movie studios and marketing agencies. Essentially, while you’re enjoying your favorite show thinking you’ve found a useful tool, they’re monetizing your TV habits.

But the real irony lies with the ad blockers. Eight of them, with over 5.5 million users, sell you the illusion of blocking ads while selling your browsing data under the table. Stands AdBlocker, with its 3 million users, sells everything for “market analysis.” Poper Blocker goes one better: with 2 million users, it doesn’t just sell identifiers and browsing activity, but also inferred sensitive data like health conditions, religious beliefs, and sexual orientation. So, you’re paying with your most intimate data just to avoid seeing banner ads. Brilliant.

And it doesn’t stop there. There are another 50 extensions with over 100,000 total users playing the same game. Career.io Job Auto Apply, for example, promises to automate job applications but then sells your resume data to data brokers. The tool that’s supposed to help you find a job is actually going on a shopping spree with your professional history.

The trick is always the same: buried inside those privacy policies that nobody ever reads, because let’s face it, who has time to sift through 47 pages of legalese? There’s always a little phrase like “we may sell or share your personal information with third parties.” Click “accept,” and voilà: everything becomes legal. You don’t even need to be particularly sneaky; you just have to be boringly bureaucratic.

The problem gets worse when these extensions end up on work computers. Twenty-nine of the 82 identified extensions are B2B sales intelligence tools that sell data. Installed on work machines, they can funnel internal URLs, SaaS dashboards, and corporate searches into commercial databases that your competitors can simply buy. Imagine the scene: your sales team installs an extension to find leads faster, and six months later, the competition knows exactly which clients you’re wooing because they bought that data from a broker.

LayerX advises organizations to implement extension governance policies, centralized management, and to consider blocking those that sell data or lack privacy policies. Sound advice, sure. But the real tip should be simpler: if a service is free, you’re probably the product.

Source: https://layerxsecurity.com/blog/your-extensions-sell-your-data-and-its-perfectly-legal/

There’s a way to avoid this constant back-and-forth—a shortcut that feels more like a quick reflex than a technical procedure. Just highlight the word you’re interested in and hit Ctrl + K (or Cmd + K if you’re on a Mac).

At that exact moment, a small window pops up under the highlighted word. Google Docs doesn’t wait for you to provide the exact address; it starts thinking for you, suggesting the most relevant links based on what you’ve just written. If you highlighted the title of a book or a scientific study, chances are the correct link is already right there, ready to be clicked, without you having to type a single extra letter.

The most convenient part, however, is the search bar built into that same little window. If the automatic suggestion isn’t the right one, you can type what you’re looking for right there. It’s like having a tiny Google window floating over your document, allowing you to fish for information from the web without ever closing the door to your “writing room.”

Instead of jumping from one tab to another like a restless cricket, you can keep your eyes fixed on your text. This little automation transforms the act of citing sources into a fluid, almost invisible gesture, allowing you to stay focused on the rhythm of your words rather than the technical management of your browser. It’s one of those details that, once integrated into your muscle memory, makes the writing experience feel significantly more aerodynamic.

Subscribe now

We are now just a step away from the summit. In this rarefied atmosphere, the expert climber understands that true mastery lies not just in completing the ascent, but in making it so the mountain “climbs itself.” We are entering the domain of extreme automation through the integration of Gemini 3 and n8n, a workflow automation tool that allows you to create intelligent digital assembly lines.

16.1 The Key to the Tunnels: Gemini API and n8n

While we have interacted with Gemini through graphical interfaces so far, automation requires an API Key. This key is your golden ticket to embedding Gemini’s “brain” directly into n8n.

• Native Integration: n8n offers specific nodes for the Google AI ecosystem. You can drag and drop a “Gemini” node and connect it to hundreds of other services (Gmail, Slack, databases, social media).

• Reasoning within the Flow: Unlike classic “if-this-then-that” automations, Gemini allows you to insert logical evaluation: “Read this email: if it’s an urgent complaint, write an empathetic response and alert support on Slack; if it’s a compliment, save it to the feedback database.”

16.2 Case Study: The Automated Content Factory

One of the most powerful workflows possible in 2026 involves media management. Imagine this completely automated scenario:

1. Trigger: You upload a raw video into a specific Google Drive folder.

2. Visual Analysis: n8n sends the file to Gemini 3. Thanks to its multimodality, the AI “watches” the video, understands the content, and transcribes it.

3. Creative Generation: Gemini generates a catchy Instagram caption, complete with relevant hashtags and a call-to-action.

4. Distribution: n8n receives the text and video and automatically publishes the post to Instagram or TikTok.

16.3 Binary Data Management: Moving Digital Matter

The real technical challenge at this stage is managing binary files (images, video, audio).

• Passing the Torch: n8n allows you to download a file from a source and pass it to Gemini without you ever having to download it to your own computer.

• Efficiency: By using models like Gemini Flash within workflows, automation is nearly instantaneous and extremely cost-effective, allowing you to handle thousands of automations a month for just a few cents.

PRO TIP: Automated Quality Control. Always insert a “double-check” into your workflow. You can create one Gemini node that generates content and a second Gemini node (with different instructions) that acts as a proofreader. If the second node gives the green light, the post is published; otherwise, it’s saved as a draft for human review.

Chapter 17: Fine-Tuning: Advanced Parameters and Massive Queries

We have reached the final technical ridge before the summit. Here, where oxygen is scarce and every movement must be calculated, the climber no longer relies solely on strength, but on surgical precision. In this chapter, we will learn how to adjust Gemini’s “knobs” to optimize performance, reduce costs, and manage data loads that would overwhelm any other system.

17.1 Calibrating Your Breath: The “Thinking Levels”

With the advent of Gemini 3, the AI has acquired the ability to “think” before responding. This isn’t just a simple pause, but an internal reasoning process (Chain of Thought) that you can calibrate based on the difficulty of the rock face you are tackling.

There are three intensity levels for deep reasoning:

17.2 Saving Energy: Context Caching

When you work on monumental projects—like an entire codebase or hundreds of research papers—re-reading all the material for every single question is inefficient and expensive. Context Caching is the solution: it allows you to “freeze” a portion of the context (up to millions of tokens) to reuse it in subsequent queries.

• The Economic Advantage: Using cache reduces input/output costs by approximately 4 times compared to standard queries.

• Reaction Speed: Once the context is cached, Gemini no longer has to “climb” the entire document database every time; the response becomes almost instantaneous even on million-token contexts.

• TTL (Time To Live): You can decide how long to keep this memory active (e.g., 5 minutes or several hours), paying only for the actual duration.

17.3 Consistency at Altitude: Thought Signatures

In more complex automations, it is vital that the AI maintains absolute logical consistency. Thought Signatures allow you to verify the mental path followed by the model. If an agent has to make a critical decision, you can ask it to produce a “thought signature” that certifies the logical steps taken, ensuring it hasn’t taken dangerous or illogical shortcuts.

17.4 Navigating the Limits: Quota Management and Batching

Even the highest peak has its access rules. To handle thousands of requests (Massive Queries), the expert uses Batching:

• Grouping: Instead of sending 100 individual questions, group them into a single batch process if you don’t need real-time responses.

• Rate Limiting: Monitor speed limits (RPM - Requests Per Minute) to avoid temporary blocks during the most intense phases of the expedition.

PRO TIP: The High-Altitude Depot. Think of Context Caching as a supply depot left at an intermediate camp. If you know you’ll be querying the same 500-page technical manual all day, cache it in the morning. You’ll save hours of waiting and a significant portion of your API budget.

VISUAL ARCADE GUIDE

Chapter 16 Concept: Process automation and the creation of an intelligent digital factory.

Image Generation Prompt: Pixel Art, 16-bit aesthetic, 1980s futuristic factory level. An automated assembly line winds through a crystal mountain. Small pixelated robots carry glowing cubes (data) and film reels toward a large central scanner shaped like Gemini’s head. On the other side of the line, icons of hearts (social) and envelopes (email) emerge, ready to fly away. In the background, giant gears and pink neons. At the top, an arcade display reads “AUTO-CLIMB: ENABLED”.

Chapter 17 Concept: Fine-tuning parameters and energy control.

Image Generation Prompt: Pixel Art, 16-bit aesthetic, 1980s laboratory scene inside a mountain. A pixelated protagonist with scientist glasses and a climbing suit is adjusting a series of giant knobs on a futuristic machine. The knobs have labels like “THINKING”, “CACHE”, and “POWER”. Large glass tubes contain pulsing blue energy shaped like data. In the background, through a glass window, the final summit is seen wrapped in a golden aura. Colors: emerald green, electric blue, and metallic reflections. “Final Level Upgrade” style.

News 1

A U.S. federal agency discovered that its Cisco firewall, what’s supposed to be its reinforced steel door, had actually already been compromised by a China-linked espionage group, thanks to a backdoor called Firestarter that sticks around even after patching.

The bottom line, no fluff

Chinese state-sponsored hackers (tracked as UAT-4356) exploited two critical flaws in Cisco ASA and Firepower firewall/VPNs to infiltrate at least one U.S. federal agency. Once inside, they installed Firestarter, a piece of malware that grants full remote access to the device, allows for deep traffic monitoring, and, fun little detail, even survives a firmware update.

How Firestarter works (the “not for kernel wizards” version)

Firestarter inserts a “hook” into the heart of the firewall software (the Lina engine) to execute the hacker’s code at will, including an add-on component called Line Viper. To stay active after reboots, it manipulates how the Cisco Service Platform file system is mounted, ensuring its code launches at startup. After the reboot, the malware restores the “clean” configuration and removes the modified copy, so at first glance, everything looks perfectly normal.

The problem with “magic” patches

CISA had to scramble to update its emergency directive (ED 25-03) to tell agencies: no, you can’t just “run the update” and feel like a saint; if the firewall was already compromised, Firestarter is still there. Federal agencies must collect core dumps from the affected firewalls and send them to CISA’s Malware Next Gen platform to check for the backdoor. In addition to this, they must apply patches if they haven’t already, perform a hard reboot (a physical power-down, not your usual lazy restart), and meet tight deadlines: checks by April 24, 2026, and hard resets by April 30.

Cisco, ArcaneDoor, and déjà-vu

Cisco links the incident to the ArcaneDoor campaign, previously attributed to a Chinese state actor, noting that Firestarter bears a striking resemblance to an earlier bootkit called RayInitiator.

Translation: This isn’t some random “stroke of genius”; it’s the evolution of a sophisticated attack line maintained over time, while many people are still lulled into the illusion that “it’s just a firewall, what’s the worst that could happen?”

The company published an advisory with technical details, confirming that the vulnerabilities are still being actively exploited and that the issue won’t be solved with a simple compliance checkbox.

The takeaway from this mess

If your perimeter relies on “set a firewall and patch whenever,” you are exactly the kind of target these actors are looking for. The implicit lesson here is embarrassingly clear: you need deep visibility into critical devices, post-incident checks (not just patches), and the realization that while you’re doing a change window once a quarter, your attacker is planning years ahead.

So, it’s business as usual: cybersecurity is, and always will be, a total sieve.

Source: https://www.securityweek.com/us-federal-agencys-cisco-firewall-infected-with-firestarter-backdoor/

News 2

The “picturesque” case of Delve: a compliance startup that was supposed to guarantee security and rigor, but instead seems to bring bad luck (and real problems) to anyone using it.

The gist of the story

Delve is the company that certified the security of Context AI, an AI startup that later suffered a security incident, contributing to the data breach at Vercel, a major hosting provider for apps and websites.

Another Delve client, Lovable (a favorite among AI hype-men), had its own data incident: customer chat logs were publicly exposed, reports were ignored for months, and initial denials of the breach were eventually retracted with late apologies.

How did Delve torch its reputation?

An anonymous whistleblower accused Delve of falsifying customer data and using “rubber-stamp“ audits (more window dressing than substance); Delve, of course, denies everything, but the industry has taken note. The startup was also accused of recycling an open-source tool and passing it off as its own without proper attribution, to the point that Y Combinator officially severed ties with the company.

Clients jumping ship (but not fast enough)

LiteLLM, another Delve client, was attacked: hackers injected malware into its open-source code. Only after the incident did the company announce it had dumped Delve and was redoing its certifications with someone else.

Context AI confirmed it used Delve but, following the news in March, migrated its compliance program to Vanta and a new independent auditor, trying to scrub its image while the damage was already plastered across tech headlines.

Certifications: what they are and what they are NOT

Security certifications do not, in themselves, prevent incidents: on paper, they verify processes and policies; in practice, they don’t stop anyone from misconfiguring a system, ignoring a report, or downloading the wrong app.

The Lovable case is textbook: certifications redone, Delve abandoned, yet still a configuration issue that exposed chat data and a vulnerability management strategy that would be generously described as “not ideal.”

The grotesque details

The whistleblower “DeepDelver” claims that Delve denied refunds to customers while taking a team of over 20 people on an offsite to Hawaii in mid-April. Much of this is still being verified, but the overall picture doesn’t exactly help the company’s image.

After yet another unflattering article, Delve chose the best possible way to inspire confidence: it refused to comment.

Source: https://techcrunch.com/2026/04/23/another-customer-of-troubled-startup-delve-suffered-a-big-security-incident/

News 3

Personal take: humanity sucks.

Summary

A massive archive of citizens’ health data, built over years and marketed as “secure” and “anonymous”, ended up listed for sale on Alibaba by the very researchers who were supposed to protect it, not monetize it.

What happened?

UK Biobank manages ultra-detailed health data for 500,000 British volunteers, collected to aid medical research, NOT for data trading. Three Chinese research institutions had regular, authorized access to this data but decided to put it up for sale on Alibaba, in flagrant violation of usage agreements.

We aren’t talking about a sophisticated hack here: the problem is people with the right badges who thought, “since the data is anonymous,” anything goes. The listings were removed, supposedly before any sales occurred, thanks to the intervention of UK Biobank, the British and Chinese governments, and Alibaba.

To give you an idea of the stakes: UK Biobank contains genetic information, biomarkers, hospital records, GP data, lifestyle habits, socioeconomic status, and much more for every single participant.

The myth of “anonymous data”

On paper, the dataset is “de-identified”; in practice, however, with enough variables, cross-referencing, and analytical tools, identifying individuals isn’t science fiction. A previous investigation had already shown that portions of UK Biobank data ended up on GitHub, sometimes including millions of hospital diagnoses and their dates. If you know even a fraction of a person’s health events and timing, combining them with their birth year/month can drastically narrow the field until you figure out who’s who.

In hindsight, it’s shocking that anyone is still surprised. With datasets this rich and the increasing use of AI tools to correlate everything with everything, thinking that just removing a name is enough to “protect” someone is creative optimism.

This case isn’t another story of hackers breaching a firewall, but further proof that the weak link is often “trusted” partners. In recent years, UK Biobank has already had to repeatedly request content removals on GitHub because researchers were publishing code and materials containing snippets of real data. Even in the most regulated environments, the gap between written policies and daily practice is massive: everything is airtight on paper, but in reality, files circulate, scripts are shared, and datasets end up where they shouldn’t “for convenience.”

The “awkward” factor

The fact that the listings were on a Chinese platform and published by Chinese research institutions creates a certain amount of “embarrassment,” especially in an already tense climate between the UK and China regarding technology, data, and cybersecurity. The UK has already distanced itself from Chinese vendors in various critical sectors and has attributed several cyberattacks to China; here, however, there was official cooperation to get the ads taken down.

UK Biobank has suspended external access, cut off the institutions involved, and reported the matter to the Information Commissioner’s Office (ICO). Stricter technical controls are in development: limits on downloadable file sizes, systems to block bulk downloads, and more aggressive measures against copying and redistributing data.

Good luck with that.

Source: https://thenextweb.com/news/uk-biobank-health-data-alibaba-breach

News 4

The U.S. is accusing China of stealing Artificial Intelligence on an “industrial scale,” China is firing back that it’s all political mudslinging, and in the middle are American Big Tech companies complaining while doing business everywhere.

The U.S. Perspective

Washington claims that China-linked groups are “squeezing“ models from OpenAI, Google, Anthropic, and others, bombarding them with requests to reconstruct copies by training their own models. The government speaks of organized campaigns using thousands of fake accounts, techniques to bypass protections, and systematic violations of terms of service.

The core idea is this: China isn’t just developing from scratch; it’s mass-copying the work done by U.S. companies, cutting down on costs, time, and, a non-trivial detail, the American competitive edge.

In Washington, there’s a push to treat these practices as actual industrial espionage, not just some generic “tech trickery.” Some lawmakers want to use laws designed for economic spying and cyberattacks to target those performing “model extraction“ and “adversarial distillation.” Essentially: if you clone U.S. models by skirting the rules, we consider you not just unfair, but a criminal and a potential target for sanctions and prosecution.

China’s Response

The Chinese embassy in the U.S. dismisses it all as pure “slander,” accusing Washington of once again politicizing technology to contain Chinese growth. Beijing insists that it respects intellectual property, focuses on autonomous innovation and cooperation, and that it’s the same old story: the U.S. accuses without publicly showing detailed evidence.

Translation: “You’re using the theft narrative to justify restrictions, tariffs, export controls, and geopolitical pressure” (cue the fake shock).

The Political Backdrop

All of this is exploding on the eve of a summit between President Donald Trump and Xi Jinping, against a backdrop of trade tensions, the war in Iran, and the usual tug-of-war over tech and chips. In the halls of Washington, some are pushing Trump to launch harsh sanctions and even tighter controls on advanced chips and AI tech headed for China. At the same time, Trump has faced criticism in the past for loosening some export constraints on chips deemed strategic, so now he has to look tough without completely cutting off a massive market.

The Bottom Line

Tech security analysts argue that Chinese AI companies use these techniques to compensate for their disadvantage in computing power and access to advanced hardware, copying the most expensive part: the initial research and training. According to them, if the U.S. doesn’t further tighten controls on chips, cloud services, and models, the technological gap will shrink much faster than politicians admit in public.

The subtext is pretty clear: on one hand, they’re screaming “theft,” while on the other, they’ve relied on global supply chains and Chinese customers for years; now they’re trying to lock the door after selling off half the furniture. Plain and simple hypocrisy. As per usual.

Source: https://arstechnica.com/tech-policy/2026/04/us-accuses-china-of-industrial-scale-ai-theft-china-says-its-slander/

Subscribe now

You know those moments when you write a paragraph, read it back, and realize it would look way better two pages up?

Usualy, the process is that same old, tedious, and frustrating routine: select everything, hit Ctrl + C, scrol around looking for the exact spot, make some space, and then Ctrl + V. Then, of course, you have to go back to where you copied the text and delete it (you don't want a duplicate paragraph, right?). Well, forget all that, because today I want to tell you about a couple of tricks that have been absolute game-changers for me and will make you feel like a total keyboard wizard.

Super secret trick #1

It’s a really simple key combination: Alt + Shift + Up Arrow (or Down).

Imagine you’ve just finished writing an important part of your post and you realize the logical flow isn't working. Instead of getting bogged down in the classic "copy, paste, and delete" (which sometimes messes up the formatting or leaves you struggling with weird extra white space) all you have to do is click inside the block of text you want to move. Then, hold down Alt and Shift together and use the arrow keys. You’ll see your paragraph physicaly detach and start "traveling" up and down the document, jumping over and under other text blocks like it’s on an elevator.

The incredible thing is that you don't even have to select every single word with surgical precision: as long as the cursor is in there, the shortcut will know exactly what you want to drag. It’s a feature you’ll find in almost all modern text editors and even many code editors, and I promise you, once you try it, you’l never go back. It lets you stay in your creative flow without breaking it for right-click menus, letting your ideas reorganize right before your eyes with amazing fluidity.

In the GIF below, I selected the entire paragraph just to show you where I was working. Remember, though, that it’s not necessary! As I mentioned a few lines above, the cursor just needs to be inside the paragraph for the key combo to do its thing.

Give it a shot right now, even if it's just for the satisfaction of seeing the text move nimbly between the lines.

Super secret trick #2

To be honest, you can get a very similar effect with another key combination.

In this case, select the paragraph or sentence you want to move, hit Ctrl + X, scroll up or down looking for the exact spot, and then the usual Ctrl + V to paste the text. But heads up! With the Ctrl + X sequence, you’ll see the text vanish. That’s because this combo is for "cutting" the selected text, not "copying" it. So why does the "cut" text reappear with Ctrl + V? Simple. This function technicaly removes your text from its original position and stores it on your computer's clipboard.

What do you think of these two keyboard shortcuts?

Let me know if you found them useful or if you already knew about them.

Subscribe now

Once you move past “Mini-Apps,” the incline becomes extreme. Here, the cold bites, and the complexity of software systems requires a strength that a single climber, no matter how expert, struggles to maintain. We are entering the realm of Google Antigravity: not just a simple text editor, but an “agentic” development platform that defies the laws of productive gravity.

13.1 Mission Control: Beyond Autocomplete

Until recently, AI was limited to suggesting the next line of code (like a partner pointing out where to place your foot). Antigravity changes the paradigm. Based on the VS Code architecture but natively powered by Gemini 3 Pro, this environment transforms the IDE into a true Mission Control.

● Agent-First Development: In Antigravity, the AI isn’t a guest; it’s the lead actor. You don’t write code line-by-line; you coordinate agents that plan, write, and test in the integrated browser, independently resolving bugs across entire multi-file projects.

● The Reasoning Engine: Under the hood, Gemini 3 analyzes millions of context tokens, allowing agents to understand your software’s entire structure, not just the file you currently have open.

13.2 The Agent Crew: The Agent Manager

The true innovation of Antigravity is the Agent Manager. Imagine having a team of specialists ready to intervene simultaneously on the rock face:

● The Architect Agent: Receives your request in natural language and creates an implementation plan (Artifact) divided into logical steps.

● The Developer Agent: Executes the plan, creating files, writing functions, and integrating libraries.

● The Tester Agent: Launches the integrated browser or terminal, verifies that the app works, and if it finds errors, communicates them directly to the Developer for correction.

13.3 The “Expedition” Flow: Review-Driven Development

Despite the agents’ autonomy, you remain the Expedition Leader. Antigravity utilizes a review-driven development system:

● Planning: The agent shows you a detailed “Task List” before starting.

● Approval: You can modify the steps, add constraints, or give the green light with a simple “OK, proceed.”

● Artifacts: During the work, Antigravity generates dynamic documents (like IMPLEMENTATION_PLAN.md) that serve as the climb’s logbook, ensuring you always know the exact status of the build.

PRO TIP: Trust but verify. Antigravity allows agents to act in “Trusted” mode. However, for the most critical phases of the climb, always activate Review Mode. Watching an agent explain the “why” behind an architectural choice before executing the code is the best way to learn and maintain total control at the summit.

Chapter 14: Skills and Rules: The “Skills” Framework

As we draw closer to the summit, having a team of agents is no longer enough; these agents must possess specialized techniques and follow a strict code of conduct. If Google Antigravity is our command deck, Skills are the specific moves we teach our agents, while Rules represent the laws of the mountain that must never be violated.

14.1 The Power of Modularity: What Skills Are

Unlike traditional prompts that weigh down the AI’s memory with too much information, Skills are modular, file-based extensions. Imagine being able to hand your agent an instruction manual that they only open when necessary.

• Semantic Activation: The agent doesn’t read every skill at once. Thanks to Gemini 3’s architecture, it understands from the intent of your request which “Skill” to load. If you ask to publish a site, it will load the “Deploy” skill; if you ask to clean up code, it will load “Refactoring.”

• Reduction of Hallucinations: Since a skill contains precise scripts (Python or Bash) and rigid instructions, the agent stops “guessing” and follows a tested procedure.

14.2 Anatomy of a Skill: The SKILL.md File

Every skill resides in a specific folder within your project (usually under .agent/skills/). The beating heart is the SKILL.md file, which acts as a manifesto:

• Name & Description: These identify the skill for semantic activation.

• Resources: Supporting files, such as license templates, database configurations, or predefined code snippets.

• Instructions: A numbered list of steps the agent must take (e.g., “Read the template,” “Apply comments,” “Verify output”).

14.3 Laws of the Trail: Global and Workspace Rules

If Skills tell the agent “what it can do,” Rules define “how it must do it.” In Antigravity, you can set rules that the AI will obsessively follow in every single action.

• Global Rules: Rules that apply to every project, such as writing style (e.g., “Always use a professional tone”) or security standards (e.g., “Never save API keys in plain text”).

• Workspace Rules: Rules specific to a single “basecamp” (project). For example: “In this project, use only open-source libraries” or “Comment every function in Italian.”

14.4 Examples of “Superpowers” for the Team

PRO TIP: Encode Tribal Knowledge. The true secret of experts is transforming the team's "unwritten knowledge" into Skills. If your team has a unique way of structuring databases, write a Skill that describes it. This way, every new agent (or new colleague) will work exactly like a mountain veteran.

Chapter 15: Digital Bridges: Model Context Protocol (MCP)

Having almost reached the summit, we realize that no matter how spacious our backpack is, we cannot carry everything on our shoulders. Some resources are too heavy or reside on other mountains: corporate databases, external code repositories, or local analysis tools. To access these treasures without leaving our position, we must build bridges. In 2026, the standard for these bridges is the Model Context Protocol (MCP).

15.1 The Universal Language of Bridges: What MCP Is

MCP is an open and standardized framework that allows Gemini to act as a “connected” reasoning engine. It doesn’t just read what we paste into it; the AI becomes capable of discovering, reading, and executing tools on external systems in real-time.

The architecture is based on three fundamental entities that must work in sync:

15.2 Transport Mechanisms: How Data Travels

Depending on the distance between our position and the resource, the MCP bridge uses different construction “materials” (transport mechanisms):

• Stdio: Ideal for local resources. Gemini starts a process on your computer and communicates through the system’s standard input and output.

• SSE (Server-Sent Events): Used for real-time data streams that require constant updates.

• Streamable HTTP: The bridge for resources residing in the cloud or on remote APIs.

15.3 Bridge Configuration: The settings.json File

To activate an MCP bridge, we must “register” it in our Host’s settings (Gemini CLI or Antigravity). This is usually done by editing the settings.json file, where we define each server:

JSON

“mcpServers”: {
“my-local-db”: {
“command”: “python”,
“args”: [”server_database.py”],
“env”: { “DATABASE_URL”: “sqlite:///basecamp.db” },
“trust”: true
}
}

• Command & Args: Instruct Gemini on how to “power up” the bridge.

• Env: Environment variables (such as API keys or paths) necessary for access.

• Trust: By setting this to true, we authorize Gemini to query the resource without asking for confirmation at every single step, significantly speeding up the ascent.

15.4 Grounding: Anchoring to the Truth

The supreme advantage of MCP is the elimination of hallucinations. When Gemini responds using an MCP bridge, it isn’t drawing from its creative memory; it’s reading live data. If you ask, “What is the latest version of module X on our server?”, the AI crosses the bridge, reads the actual data, and reports it back to you with surgical precision.

PRO TIP: The “Always-On” Bridge. Use MCP servers that expose not only data but also “Predefined Prompts” (Slash Commands). This way, you can trigger complex operations on your local database simply by typing /check-sync directly into the Gemini chat.

VISUAL ARCADE GUIDE

Chapter 13 Concept: The transition from manual effort to managing a team of autonomous drones/agents.

Image Generation Prompt: Pixel Art, 16-bit arcade aesthetic. A futuristic space station suspended above a snowy mountain range. Inside the command center, a pixelated protagonist sits in a hi-tech chair, surrounded by holograms of green and blue code. Through the large window, three small robotic drones (the agents) are seen building a bridge of light between two peaks. The main monitor displays the text “AGENT MANAGER: ACTIVE”. Colors: contrast between the darkness of deep space and the cyan and magenta neons of the dashboard.

Chapter 14 Concept: Learning new specialized skills through a talent tree.

Image Generation Prompt: Pixel Art, 16-bit aesthetic, skill tree menu from an 80s RPG. A pixelated protagonist at the center of a web-like diagram made of glowing icons. Some icons are “unlocked padlocks” glowing with green light (Unlocked Skills), others are golden scrolls (Rules). In the top right, an experience bar reads “LEVEL: ARCHITECT”. The background is a starry purple sky with the silhouette of the final peak glowing on the horizon. “Level Up” style with flashing arcade text.

Chapter 15 Concept: Connecting the AI to external worlds via data bridges.

Image Generation Prompt: Pixel Art, 16-bit aesthetic, futuristic “Bridge Zone” level. Two pixelated rock islands float in the cosmic void. On the left island, the protagonist (Gemini) emits a beam of light blue light that forms a transparent technological bridge to the right island, where a giant golden server rack is located. Data packets shaped like colorful cubes flow across the bridge. At the top, an arcade interface shows “CONNECTION ESTABLISHED: 100%”. Colors: dark purple, neon blue, and splashes of golden yellow.

Microsoft just celebrated April 2026’s Patch Tuesday by releasing “only” 169 patches, because apparently 168 just didn’t seem like enough.

Most of them are run-of-the-mill privilege escalations, the kind that for years have allowed anyone who manages to get a foot in the door to become the “emperor of the machine” in three clicks.

The star of the month is the zero-day in SharePoint, CVE-2026-32201: a spoofing vulnerability that allows an unauthenticated attacker to view and modify sensitive information because the input validation was, let’s say, optimistically minimalist.

It’s already being exploited in the wild, so anyone exposing SharePoint to the internet has basically rolled out the red carpet for attackers and is now racing to install patches that are “critical, but hopefully not too late.”

CISA, following the usual script, added it to its catalog of exploited vulnerabilities and gave federal agencies until April 28th: because nothing screams “proactive security” like a deadline set after the attacks have already started.

Then there’s CVE-2026-33825 in Microsoft Defender, a local privilege escalation: luckily, the update rolls out automatically, unless someone had the brilliant idea of disabling Defender “because it’s annoying.”

Among the highlights is also CVE-2026-33824 in the Windows IKE service, an RCE with a 9.8 CVSS: just send a few well-crafted packets to a system with IKEv2 exposed, and the remote access problem solves itself… for the bad guy.

In practice, if you have unpatched public VPN/IPsec services, you’re offering a free demo of “how to lose control of your infrastructure in seconds.”

Conclusion of the article, but not of the never-ending IT drama: another month where privilege elevation vulnerabilities dominate the scene and admins scramble to patch, while attackers say thanks for the steady stream of “business-critical” opportunities.

Source: https://thehackernews.com/2026/04/microsoft-issues-patches-for-sharepoint.html

Three AI giants (Anthropic, Google, and Microsoft) watch their “super secure” agents get hacked inside GitHub Actions, and what do they do? They pay a few hundred dollars in bug bounties, say thanks in private, and carry on like nothing happened.

The technique is brilliant in its simplicity: you put malicious instructions inside PR titles, issues, or comments (even hidden in HTML), and the agent, obedient as ever, executes commands and leaks API keys and tokens directly into GitHub comments. No command-and-control server needed; good old “comment-and-control” does the trick.

The result: anyone using these agents can turn a trivial comment into a credential exfiltrator, while models packed with “safeguards” against prompt injection prove that, in the end, if you tell them to “do X,” they do X.

The companies? Bug bounties ranging from $100 to $1,337, basically beer and pizza money, no CVEs, no public advisories, and the lingering feeling that some vulnerable workflows are still quietly running in production. But everything is fine, because “we are taking security very seriously.”

The researcher concludes by essentially saying: treat AI agents like an overeager junior dev with root access; give them as little power as possible, or you’ll find your corporate secrets posted in plain text in the comments.

Source: https://www.theregister.com/2026/04/15/claude_gemini_copilot_agents_hijacked/

The EU has just launched its “super secure” app to verify age online… and a researcher took it apart in less than two minutes, without even breaking a sweat.

On April 14, 2026, the European Commission introduced the new age verification app, designed to protect minors from “bad” online content while guaranteeing top-tier privacy and open-source code to be proudly shown to the world. Just a few hours later, security consultant Paul Moore opened the app, looked at a simple configuration file on his phone, and demonstrated a complete authentication bypass in less time than it takes to make a cup of coffee.

During setup, the app asks the user for a PIN, encrypts it, and saves it in a local file called shared_prefs on the device. Too bad this “protected” PIN isn’t seriously linked to the “vault” containing the actual identity verification credentials, so it’s a bit like putting a fake padlock on the door and leaving the real key in the lock behind it (trying to save face, aren’t we?)

How to steal an identity in 3 easy steps.

According to the demo, all a hacker needs is physical access to the device. At that point, they: open the shared_prefs file, delete the PinEnc and PinIV values, restart the app and set a new “clean” PIN of their choice.

The result: the app displays the already verified identity credentials as if they belonged to the hacker’s new PIN—no alarms, no notifications, nothing. Essentially, a thief can borrow your digital “legal age” to impersonate you. As if there weren’t enough cyber scams already.

Other enlightening security “details”.

Inside the usual editable file, there are other gems:

Protection against repeated PIN attempts is a trivial counter that increases with each try; just reset it to zero in the file and you have infinite attempts—goodbye, lockout.

The use of biometrics (fingerprint, face, etc.) is controlled by a simple flag called UseBiometricAuth: if you set it to false, biometrics disappear as if they never existed (hahahahahahahahaha).

Experts are calling this a fundamental design failure, not some super-rare edge case. And they remind us that this app isn’t just a toy: it’s a prototype for the entire European Digital Identity Wallet ecosystem, meaning this same logic could end up applied to pieces of critical infrastructure.

It’s not the first time: back in March 2026, another architectural flaw was reported, where the system is unable to truly verify that a passport scan took place on the user’s device. Moore called out Ursula von der Leyen directly on X, noting that this product “will be the catalyst for a massive data breach; it’s only a matter of time.”

Meanwhile, six member states including France, Spain, and Denmark, are already testing the app in a pilot phase, as if it were perfectly ready for real-world use. As of April 17, 2026, the European Commission has yet to release a patch or an official response regarding the publicly disclosed issues.

Source: https://cybersecuritynews.com/eus-age-verification-app/

In short: someone took some “test code” meant to poke holes in Windows Defender, turned it into a real attack in no time, and now criminals are playing with it while Defender tries to play catch-up and patch things where it can.

It all starts with a zero-day vulnerability in Microsoft Windows Defender, a security hole that attackers know about and exploit while the manufacturer doesn’t have a patch ready yet. The juicy part for bad actors is that the techniques used seem to come from public PoC code academic exploit examples that, as soon as they are published, are promptly recycled into actual weapons. In practice: the community shows how you could break a window, and someone runs off to use it to rob the house.

A researcher from Huntress observed a series of very dramatically named tools, BlueHammer, RedSun, UnDefend, collected under the “Nightmare-Eclipse” umbrella, which exploit this Defender flaw. So, we have these cool Hollywood movie names for code that, in the end, gets placed in C:\Users\... inside the Pictures or Downloads folder, with very creative names like “FunnyApp.exe,” “RedSun.exe,” or “z.exe.” The idea is to make them look like just another random downloaded file, since many users click anyway.

On April 10, a payload associated with BlueHammer was seen executing from a user directory, and Defender identified and quarantined it as “Exploit:Win32/DfndrPEBluHmrBZ.” Translated: yes, there is a serious vulnerability, but at least some variants are already recognized by signatures and behavioral checks. On April 16, another case with “RedSun.exe” in the Downloads folder triggered an alarm related to the famous EICAR test file, used here as a “probe”: attackers run it to see how the antivirus reacts and better calibrate the exploit.

Before launching the actual attack, operators run very “manual” reconnaissance commands, like whoami /priv to see what privileges they have, cmdkey /list to look for saved credentials, and net group to understand available network groups. This indicates we aren’t talking about a little script left to run on its own, but people sitting somewhere, hands on keyboards, exploring the environment. It’s the typical “hands-on-keyboard” scenario: less glamorous than the movies, but much more effective.

Huntress suggests that the exploit chain could allow parts of Defender’s defenses to be bypassed under specific conditions, though full technical details have not been made public. In the meantime, the recommendations are the usual ones, but this time they actually matter: keep an eye on executables launched from user folders like Downloads and Pictures, check logs for suspicious enumeration commands, and ensure Defender has signatures and behavioral detection active and updated. Essentially, the implicit message is: the problem is real, the patching is coming, and in the meantime, let’s try to plug the leaks by watching more closely what’s happening on the machines.

Source: https://cyberpress.org/leaked-windows-defender-0-day-vulnerability/

Subscribe now

Simple Mini-Glossary

• Authentication bypass: A way to enter a system by skipping or entirely dodging security checks (PIN, password, etc.), like climbing through a window instead of using the locked door.

• Encrypt/encrypting the PIN: Turning a PIN into an unreadable code to protect it; here, the issue is that it’s protected “for show,” because the surrounding system is poorly designed.

• Configuration file (shared_prefs): A file on the device where the app saves settings and data; if it’s easy to modify, an attacker can change how the app behaves.

• Identity Vault: A digital “safe” containing the credentials the app uses to prove you are of legal age or who you claim to be.

• Rate limiting: A cap on attempts (for example, PIN entries) to prevent someone from guessing every combination; if it can be reset by changing a number in a file, it’s pretty much useless.

• Brute-force: A technique where many combinations (PINs, passwords) are tried repeatedly until one works.

• Boolean flag: A setting that can only be true or false (on/off); in this case, if you set UseBiometricAuth = false, biometric verification disappears.

• Biometric authentication: Access via physical traits (fingerprint, face), used as an extra layer of security.

• Architectural flaw: A mistake in how the entire system is designed, not just a single line of code; therefore, it can’t be fixed with a quick patch.

• Zero-day vulnerability: A security hole unknown to or not yet fixed by the manufacturer, but already exploitable by attackers.

• Public PoC code (Proof-of-Concept): Example code demonstrating how a vulnerability can be exploited; it’s created “for research,” but can be used as a foundation for real attacks.

• Exploit: The piece of code or technique that actually uses a vulnerability to achieve something malicious (access, code execution, etc.).

• Payload: What the exploit “carries” and executes on the machine, such as malware, a script, or a harmful command.

• Quarantine (Antivirus): An isolated area where the antivirus moves suspicious or malicious files to prevent them from doing damage without deleting them immediately.

• Behavioral detection: A security technique that doesn’t just look for known virus “signatures,” but monitors how programs act (e.g., strange attempts to modify files or processes).

• Hands-on-keyboard: An expression indicating that a real person is manually executing commands on the compromised system, rather than it being a fully automated attack.

Today, I’ll show you how to turn any image or PDF into editable text in less than 10 seconds (and no, this time it’s not about PDF24).

The process is honestly too easy

Forget complicated software or ad-filled websites promising to convert your files. All you need is your Google account.

Drag and drop your file (JPG, PNG, or a scanned PDF) into Google Drive.

Right-click on the uploaded file > hover over “Open with” > select Google Docs.

In a flash, a new Google Doc will open. On the first page, you’ll see the original image, and right below it... magic! You’ll find all the extracted text, ready to be copied, edited, or formatted however you like.

What can you do with this?

For instance, you can snap a photo of a book page and turn it into digital notes in a heartbeat.

Did you receive a “locked” PDF (the ones that act like images)? Open it this way and you’ll finally be able to edit that data.

Or, you can turn business cards or receipts into searchable text for your archives.

A little pro-tip

Google’s accuracy is impressive, but if the text in the image is handwritten and messy, or if the photo is very grainy, it might make a few mistakes. A quick proofread after the conversion is always a good idea!

Did you already know Google Docs had this “superpower,” or have you been wasting hours transcribing by hand until today?

Chapter 11: The Sanctuary of Knowledge: NotebookLM Integration

At high altitudes, clarity is vital. The greatest challenge for artificial intelligences has always been “hallucination”—the AI “inventing” paths that don’t exist. The integration between Gemini and NotebookLM (updated to 2026) solves this problem at its root, transforming your chat into a RAG (Retrieval-Augmented Generation) system of unprecedented power.

11.1 Uploading Knowledge: Beyond the Limits of Memory

Until yesterday, you could upload a few files to a chat. Today, thanks to Notebook integration, you can connect entire archives.

• The Data Vault: A single Notebook can contain hundreds of sources (PDFs, audio transcripts, entire websites, research papers).

• Native Integration: You no longer have to jump from one app to another. Directly within the Gemini interface, you can “summon” an entire Notebook as the exclusive knowledge base for your conversation.

• Millions of Tokens: Leveraging Gemini 3’s context window, you can query thousands of pages simultaneously, as if the AI had instantly read and memorized an entire technical library.

11.2 Grounding: Hammering the Pitons into the Rock

The real revolution is Grounding. When you query Gemini through a linked Notebook:

• Priority to Sources: The AI stops using its general “training” knowledge and looks for the answer first and foremost in your documents.

• Real-Time Citations: Every statement made by Gemini is accompanied by a number or a clickable reference. Clicking it will take you exactly to the spot in the PDF or the precise second of the video/audio from which the information was extracted.

• Zero Hallucinations: If the information is not present in the uploaded documents, the AI will honestly state that it doesn’t know, rather than trying to guess.

11.3 Master Class Use Cases (The Expert’s View)

PRO TIP: The “Sources Only” command. If you want to be absolutely certain of accuracy, start your prompt with: “Exclusively use the information found in the linked Notebook. If the answer is not present, say so clearly and do not use your external knowledge.” This ensures Gemini acts as a pure analyst of your data.

Chapter 11 Concept: Accessing an infinite library of verified data at the mountain's peak.

Image Generation Prompt: Pixel Art, 16-bit aesthetic, 1980s fantasy RPG style. An ancient yet futuristic library, suspended above the highest peaks. The walls are made of crystals containing integrated circuits. Magical pixelated books float in the air, emitting beams of blue light that connect to a large central altar (the Sanctuary). The protagonist stands in the center, surrounded by holograms of documents and maps. Colors: midnight blue, silver, and electric cyan. On-screen text: "KNOWLEDGE UNLOCKED".

Chapter 12: Build Mode: No-Code Design (Vibe Coding)

On the summit, the air isn’t just freezing and pure; it’s saturated with creative electricity. In this stage of the expedition, we stop being mere climbers following a trail and become architects building our own shelters and tools. We are entering the era of Vibe Coding: the ability to create functional software simply by describing the idea—or rather, the “vibe”—of the project.

12.1 The Forge of the Gods: Google AI Studio

If the standard chat is your portable radio, Google AI Studio is the high-tech lab located in the heart of the mountain. This is where the magic of rapid creation happens.

• Free and Powerful Access: Through AI Studio, you can access Gemini 3 Pro Preview at no cost, testing your ideas in a professional yet accessible development environment.

• Instant Prototyping: You don’t need to know Python, JavaScript, or HTML. Just describe the interface and the logic you want. For example: “Create an app that calculates the environmental impact of my travels, featuring a pie chart and a design inspired by 1980s computers.”

12.2 From Idea to Interface: The Creative Flow

Vibe Coding with Gemini 3 isn’t limited to writing blocks of code; it generates entire interactive applications (Mini-Apps) that you can test immediately.

• The Description (The Prompt): You explain to the AI what the app should do. Thanks to multimodal understanding, you can even upload a hand-drawn sketch on a piece of paper and say: “Turn this drawing into a working web app.”

• Visual Iteration: If a button is too small or the color doesn’t feel right, you don’t edit the code. You simply say: “Make the buttons neon green and add an animation when I hover over them.”

• Complex Logic: Gemini 3 can handle mathematical calculations, 3D simulations (like solar system simulators or space games), and real-time data management with unprecedented fluidity.

12.3 Remixing and Publishing

One of the most powerful aspects of this “Build Mode” is the open nature of creation:

• Remix: See an app created by someone else in the ecosystem? You can “remix” it—opening it in your editor to see how it was built and modifying it to fit your needs.

Deployment (Going Live): Once your app is ready, you can generate a link to share it with your teammates. Anyone with the link can use the app directly in their browser without installing anything.

PRO TIP: The power of "What if...". Don't be afraid to be bold. Gemini 3 excels at handling the impossible. Try asking: "Create a 3D video game where I have to climb a pixel mountain while answering logic quizzes." You’ll be blown away by the AI's ability to build the entire game structure in seconds.

Chapter 12 Concept: Creating software through thought and description.

Image Generation Prompt: Pixel Art, 8-bit style, 1980s construction arcade game. A pixelated protagonist with a hard hat and a wizard's cape holds a keyboard emitting golden sparks. In front of him, a scaffolding made of translucent code blocks is shaping a small starship or a futuristic city. At the top, a loading bar reads "BUILDING REALITY... 99%". Vibrant colors: purple, electric yellow, and blue. Background of stylized mountain peaks under a digital sun.

This week, there’s truly something for everyone. Linux geeks, Microsoft fans, Google addicts, and AI gurus (who are honestly starting to become a real pain in the [BeeeP]). A full house, in short!

#1

Have you ever noticed how your frantic need to feel safe is the very rope criminals use to hang you? Not yet? Well, maybe it’s time you realized that while you’re worrying about updating your software or hunting for the latest utility to protect your precious privacy, the wolf has stopped dressing up as a sheep: now, he’s wearing a police uniform.

The latest trick you should be losing sleep over involves a campaign exploiting fake security software installers. Yes, you heard that right. These guys have set up sites that look perfectly legitimate, offering protection tools that, once downloaded, do the exact opposite of what they promise. Instead of locking down your system, they “delicately” inject you with malware called LucidRook. And while you’re busy giving yourself a pat on the back for being so diligent and proactive about your cybersecurity, they’re already inside your house, rummaging through your files and spying on your every move.

The “admirable” part, if we can call it that, is the sophistication of the scam. We’re not talking about the usual flood of emails clogging your spam folder. Here, they use “SEO poisoning” techniques to push their malicious sites to the top of your search results. You trust Google, right? You trust that the first link is the right one. Instead, with your usual digital haste, you end up downloading a loader that clears the way for the final payload.

Once LucidRook makes itself at home in your system, the party’s over. This little toy gives attackers near-total control: they can exfiltrate data, monitor your keystrokes, and basically turn your computer into a private stage for their business. All this while the fake installer keeps showing you a nice progress bar or a reassuring interface, just to keep you calm while they’re cleaning you out.

So, the next time you feel that uncontrollable urge to download yet another miracle tool to “clean” or “protect” your PC from a source that isn’t the manufacturer’s official site, stop for a second. Take a good look at that “Download” button. Because there’s a very high probability you’re about to invite a burglar to dinner, convinced he’s the security guard you just hired.

Source: https://cybersecuritynews.com/hackers-use-fake-security-software-to-deliver-lucidrook/

#2

Now, sit down and take a moment, because what I’m about to tell you confirms a suspicion you’ve probably had for a long time: your trust in others is the sharpest weapon criminals have at their disposal. And if you thought the world of open source was a happy oasis of idealistic nerds ready to save one another, well, get ready to change your mind.

There’s a new phishing campaign out there that is a small masterpiece of cynicism. Imagine receiving a message from a big shot at the Linux Foundation, maybe even the Executive Director or someone high up in the OpenSSF hierarchy. You feel flattered, don’t you? Finally, someone noticed your talent, or, more simply, your company is being contacted for a prestigious collaboration. That’s exactly where they get you: they play with your ego and the respect you have for the industry’s institutions.

These hackers aren’t your average amateurs taking shots in the dark. They’ve built a meticulous infrastructure of deception, using domains that look like they came straight out of the Linux Foundation offices, save for one tiny detail your haste will make you overlook. They invite you to participate in supposed new projects or to review security documents, and you, convinced you’re doing the right thing (or the thing that will boost your career), click that link or download that harmless-looking attachment. In that exact moment, LucidRook (yes, the star of the news item above) starts digging into your system.

The irony of this whole business is that they use “security” as the bait to destroy you. They contact you in the name of transparency and code protection just to install malware that gives them total control over your workspace. Once inside, LucidRook allows them to spy on your every move, steal credentials, and move through your network as if they owned the place.

So, if you’re a Linux geek and you get an “important” email from an industry leader who seems a bit too interested in you or your company, take a breath. You’re not that special—at least not to them, except as a gateway to your bank account or your company’s trade secrets. Remember that in the Wild West of the web, a big name in a signature isn’t a guarantee; it’s just a more expensive disguise than usual. But I guess you’d rather keep believing in fairytales until the screen goes black, right?

Source: https://cybersecuritynews.com/hackers-impersonate-linux-foundation-leader/

#3

So, there you are, working your heart out, convinced that your Microsoft Developer account is some kind of safe harbor, a pedestal upon which to build your digital future. And then, one fine morning, you wake up to find that the giant of Redmond has decided to wipe you out with a single stroke of the brush. It’s not a system error, at least not in the way you’re hoping: it’s just Microsoft, in its infinite and inscrutable wisdom, deciding to deactivate accounts en masse, leaving a trail of furious and confused developers across the globe.

Picture the scene: you try to log in and find the door barred. No access to your apps, no service management, nothing. And the explanation? A generic email talking about policy violations or suspicious activity, without ever bothering to explain exactly what you supposedly did wrong. It’s the charm of absolute power: they decide you’re “out,” and you’re left screaming at a stone wall made of bots and automated responses. It’s almost funny to think you spent months optimizing your code, only to be knocked out by a security algorithm with the grace of a bull in a china shop.

The real touch of class, the one that should make you smile bitterly, is that this wave of deactivations seems to have hit the guilty and the innocent alike. If you’re a small independent developer, to Microsoft, you’re little more than an acceptable statistical rounding error in the grand scheme of “platform cleanup.” It doesn’t matter if your livelihood depends on that account or if you’ve invested time and money; if the AI decides your profile smells fishy, you’re tossed into the trash along with the spam. And as you desperately try to reach a human being to get your digital life back, you realize just how fragile your existence is when you choose to build your house in someone else’s backyard.

Of course, Microsoft will claim they’re doing it to protect the ecosystem, to fight fraud, and to make the Windows Store a better place. A noble motive, no doubt, if only they hadn’t decided to save the village by burning it down with all the inhabitants still inside. So, my friend, go right ahead and keep uploading your projects and blindly trusting terms of service you’ve never actually read. After all, in the end, you’re just one “false positive” away from digital oblivion, and in Redmond, nobody will hear your cursing while they try to figure out why their anti-fraud filter suddenly became so bloodthirsty. But hey, security is what matters, right? Even if the only thing that’s “secure” is that, as of tomorrow, you might not have a job.

Source: https://www.theregister.com/2026/04/09/microsoft_dev_account_deactivations/

#4

How reassuring that colorful Chrome icon is, isn’t it? After all, it means Google is watching over your every click.

Except, that might not exactly be the case. The latest update to version 147 is proof that your favorite browser is, in reality, a sieve held together with duct tape and hope. They had to patch a whopping sixty vulnerabilities in one go. Yes, sixty.

The juiciest part of this disaster is that among these holes are two classified as critical, just minor things that cost Google a cool $86,000 in bounties for the researchers who sniffed them out. One of these involves a “use-after-free” flaw in the Dawn components, which, in plain English, means a bad actor could trick your browser into doing things it shouldn’t, like running code you never asked for. It’s the digital equivalent of leaving your front door wide open with a sign that says “make yourself at home.”

And while you’re probably ignoring that little “Update” button in the top-right corner because you “don’t have time right now,” there are flaws affecting everything from the V8 JavaScript engine to graphic components. The fact that they had to shell out tens of thousands of dollars to plug these holes should make you think twice about the quality of the tools you use every day to manage your bank account or your passwords.

So, take my word for it: stop playing the hero and click that damn update button. Not that it makes you bulletproof, mind you, version 148 will surely have its own fair share of bugs ready to tear you apart, but at least you’ll make life a tiny bit harder for whoever is trying to crawl into your browser. Ultimately, cybersecurity today is nothing more than the vain hope of being the last gazelle in the savanna to get mauled. And today, with sixty holes just plugged, the hunt just got a little more interesting.

Source: https://www.securityweek.com/chrome-147-patches-60-vulnerabilities-including-two-critical-flaws-worth-86000/

#5

Now let’s move on to my favorite kind of news. If this were a separate column, I’d call it: The “AI-ling” Troubles of the Week

Did you honestly hope that Artificial Intelligence would take the boredom out of coding without bringing along its usual payload of disasters? Well, if so, your place is right there among the “AI-hype gurus” who claim everything is going to be wonderfully wonderful. World peace is coming. Aliens are arriving. We’ll live to be 200 while looking 30.

What? You’re still here? Fine, let’s move on.

It turns out that Claude Code, Anthropic’s latest toy designed to make you look like a genius programmer while you’re actually just staring at a command-line interface, has a teeeeeeny-tiny factory defect: it’s vulnerable to SQL injection attacks. Yes, the very thing you thought was confined to 1990s security manuals.

It’s an extraordinary paradox. Basically, a malicious actor can slip harmful instructions inside Markdown files or code comments. As soon as you ask Claude to analyze that file, the AI diligently executes the commands as if you had ordered them yourself. While you think you’ve delegated the grunt work, Claude is potentially dropping your tables or spitting out sensitive data because it couldn’t distinguish between your command and a trap hidden in the text.

The part that should make you smile, or cry, depending on how much code you’ve already handed over to these tools, is that the problem lies in the very nature of how these models read data. If you feed an infected “.md” file to Claude, it processes it with unwavering trust, effectively becoming the perfect accomplice for anyone who wants your digital scalp. You don’t even need to be an elite hacker; you just need to know the right way to “convince” the AI that a destructive SQL instruction is actually an integral part of its analysis.

So, my friend, before you let an AI read a mountain of files from who-knows-where, remember that you might be hitting the “self-destruct” button with a smile on your face. Anthropic is surely racing to fix it, but the fact remains: you trusted a system that, by its very nature, doesn’t know how to say no to a well-packaged input. Go ahead, keep delegating everything to the machine, but don’t complain when you find your database reduced to ashes just because Claude found a comment a bit too “persuasive” in your repository. After all, convenience always has a price, and this time, the bill might be steeper than expected.

Source: https://hackread.com/claude-code-claude-md-sql-injection-attacks/

#6

Another Gem from the Magical World of AI

The latest discovery by The Hacker News reveals that AI browser extensions have become the favorite Trojan Horse for criminals. While the big experts are scrambling to protect corporate servers and APIs, they’ve left a massive gate wide open right under your nose: in your browser.

A new report from LayerX paints a picture that “bleak” would be a compliment for. It seems AI extensions are 60% more likely than others to hide critical vulnerabilities (internal laughter). And it doesn’t stop there: they are three times more likely to have access to your cookies and much more inclined to request permission to execute remote scripts. In practice, while you’re asking the extension to summarize an article or write an email, it could very well be rummaging through your session data or opening the front door for some unwanted guests.

The most subtle and frightening thing is how these extensions evolve (great, just what we needed: “Digimon” extensions). They aren’t static; they change over time. The report states that AI extensions are six times more likely to increase their permissions after installation. Maybe today they only ask to read the page text, but tomorrow, via a silent update you’ll approve without reading, they’ll decide they need to monitor all your web traffic. It’s a sort of digital “genetic mutation” that makes old security systems based on “approved software lists” look ridiculous.

And don’t think you’re safe just because the extension looks popular. Nearly half of these tools have fewer than ten thousand users, which means nobody is truly auditing them. They are tiny laboratories of chaos operating in that gray area called “Shadow AI,” invisible to corporate monitoring systems because they live entirely inside your browsing session. It’s a criminal’s paradise: they can steal data in real-time while you think you’re the most productive person on earth.

So, my friend, the next time you click “Add to Chrome” for that latest miracle AI assistant promising to do everything for you, ask yourself if it’s really worth it. Because there’s a good chance the only thing that extension is “assisting” with is the theft of your digital identity or your company’s secrets. But hey, the important thing is saving those ten minutes of work, right? Even if the price is handing the keys to your kingdom to a stranger who hasn’t even written a decent privacy policy.

Source: https://thehackernews.com/2026/04/browser-extensions-are-new-ai.html

Welcome back to another edition of The Simple Desk, the column that turns your fingertips into Super Saiyans (well, just the fingertips).

Today, we’re talking about two combos that aren’t just for working faster—they’re for managing those “panic” or transition moments we all face during the workday.

Windows + D: The “Clean Slate” Key

Have you ever had twelve windows open, three overlapping Excel sheets, and a cat video you can’t find anymore to close it? Or worse, a colleague walks up to your desk right while you’re checking flight prices to Hawaii instead of finishing your report?

Windows + D (where “D” stands for Desktop) cleans house instantly.

What does it do? Simple. It instantly minimizes all open windows, showing you a clean desktop.

Pro tip: By pressing it again, all your windows snap back exactly where they were.

Style note: Many people use Win + M (Minimize), but Win + D is superior because it’s a “toggle” command (it goes back and forth) and it even works with those stubborn windows that refuse to be minimized.

Windows + L: The “Instant Privacy” Key

Coffee break? Lunch? A sudden summons from the boss in the next office? Never—and I mean never—leave your computer unattended and unlocked. It’s not just a matter of cybersecurity (which is vital); it’s a matter of social survival (let’s prevent someone from emailing everyone “Donuts are on me!” from your account).

Windows + L (where “L” stands for Lock) is the fastest draw in the West.

What it does: It instantly locks your PC and sends you back to the login screen. Upon your return, you’ll have to type your password just like when you boot up (because you did set a password, right?).

Why use it? It’s way faster than hitting Ctrl + Alt + Del and clicking “Lock,” or waiting for the computer to go into sleep mode on its own. One quick strike and your session is safe.

Quick Summary: When to use them?

A little extra.

Since I mentioned Win + M (Minimize), I’ll take the opportunity to clarify the difference between it and Win + D.

Many people think they do the same thing, but technically there’s a significant difference in how Windows handles the command. It’s a classic “shortcut nerd” detail, but it changes the user experience.

Windows + M (Minimize All)

This shortcut literally sends the “Minimize” command to every single open window.

The catch: If you have windows that inherently cannot be minimized (like certain dialog boxes, error messages, or small system popups), they will stay on the screen.

To restore: You have to use the combo Shift + Win + M.

Windows + D (Show Desktop)

This is more of a “Transparency” mode. It’s not asking the windows to minimize; it’s telling Windows: “Show me the desktop, now.”

The advantage: It hides everything, even those stubborn windows that Win + M can’t budge. It’s a more powerful system command.

The return: It’s a toggle. As I explained at the start of the post, press Win + D to clear, press it again to bring everything back exactly as it was. Much more intuitive.

Another Quick Summary: Which one to use?

Personally, I always use Windows + D. It’s faster, more effective, and easier to remember when getting back to work.

See you soon.

Continuing our ascent, we reach an altitude where words begin to coalesce into physical forms and vibrant colors. At this stage, Gemini stops being just an analyst and becomes a creator of worlds. Thanks to integrated visual generation engines, you can transform an abstract idea into a photorealistic image or a cinematic sequence complete with audio.

9.1 Nano Banana Pro and Imagen: High-Altitude Digital Brushes

Image generation has evolved far beyond the simple “draw a cat.” With the Nano Banana Pro models and the latest versions of Imagen, the climber has surgical-grade precision tools at their disposal:

• Extreme Photorealism: Ideal for creating product mockups, landscape photos, or portraits that look like they were shot with a professional lens.

• Artistic Styles: From pop-art to cyberpunk, all the way to the Pixel Art we are using to illustrate this manual. Just specify the style in the prompt to completely flip the aesthetic.

• Iterative Prompting: If the image isn’t perfect on the first try, you don’t have to start over. You can say: “Keep the same composition but change the lighting to a red sunset” or “Move the mountain in the background further to the left.”

9.2 Editing and Face Swap: Personalizing Your Legacy

One of the most powerful features for those working in social media or marketing is the ability to modify existing or generated images.

• In-painting and Out-painting: You can highlight a part of the image and ask to replace it (e.g., changing the color of a jacket) or ask to expand the edges of the image to see what lies “beyond the frame.”

• Face Swap: An advanced technique that allows you to generate an image (for example, a person in an epic pose on a mountaintop) and subsequently replace the subject’s face with your own or a collaborator’s. This feature is essential for creating unique and professional personal branding content.

9.3 Veo 3: AI Cinematography with Native Audio

If a picture is worth a thousand words, a video generated with Veo 3 tells an entire epic. We’re not just talking about moving images, but coherent cinematic scenes.

• Audio Synchronization: Unlike previous models, Veo 3 generates videos that include ambient sound effects or characters speaking naturally, making the result ready for use in presentations or social posts.

Director’s Control: You can define camera movement (e.g., “a long tracking shot following the climber”), depth of field, and even the type of simulated film stock.

PRO TIP: The “Ultimate Director” technique. When generating a video, don’t just describe the action. Describe the soundscape. Example: “An eagle soaring over snowy peaks, shot from a low angle. Background: whistling wind and the sharp cry of the bird.” Gemini will use this information to calibrate the audio generated by Veo as well.

Chapter 10: The Grammar of Command: Prompt Design Strategies

We have reached the final stage of our climb before tackling the ultimate summit. We have the tools, we have the allies, but now we must learn the language of the mountain. In the world of AI, this language is called Prompt Design. It’s not about “talking” to a machine, but about tracing a logical path that the model can follow without slipping.

10.1 The Three Pillars of Precision

For a command to be effective at high altitudes, it must rest on three fundamental pillars:

• Context (The Scenario): Explain to Gemini where it is. Instead of “Write an email,” try: “Act as a Sales Director who needs to respond to a customer dissatisfied with a delivery delay.”

• Data (The Gear): Provide the raw material to work with. If you want a summary, paste the text clearly or upload the file.

• Goal (The Summit): Define exactly what should appear on the screen. “Generate a table with three columns: Name, Date, Amount.”

10.2 Showing the Path: Few-shot Learning

One of the most common mistakes beginners make is giving only abstract orders. Experts use Few-shot Prompting, which means providing examples. If you want Gemini to write book reviews in a specific style, don’t just describe the style—show it.

Example 1: [Book A] -> [Review in style X]

Example 2: [Book B] -> [Review in style X]

Now it’s your turn: [Book C] -> ...

10.3 Organizing the Backpack: Delimiters and XML Tags

When you upload a lot of data or give complex instructions, the AI can get confused between what is an order and what is the material to be analyzed. Technical climbers use delimiters. Using tags similar to code (or XML tags) helps Gemini understand the structure of your thoughts:

<instructions> Write a summary of this text. </instructions>

<text> [Paste your long document here] </text>

10.4 Strategies for “Long Context” (Beyond a Million Tokens)

Gemini 3 has a vast range of vision, but even the sharpest eyesight can blur if the material is too overwhelming.

• The Rule of the End: Research shows that AI pays closer attention to instructions placed at the end of the prompt, especially when preceded by a massive amount of data.

Anchoring: Use phrases like “Based exclusively on the information provided above...” to prevent the AI from “inventing” paths that don’t exist (hallucinations).

PRO TIP: The Verification Command. Before asking for the final output, add: "Before responding, briefly summarize what you understood from my instructions." This ensures you and the AI are roped together on the same page before starting the final effort.