Mark Farragher’s Post

“…An AI-assisted campaign is spreading more than 300 poisoned packages for diverse assets ranging from developer tools to game cheats…” https://lnkd.in/g82xf9x5

What if a single AI model just changed the security landscape overnight? That 181 vs 2 exploit success rate between Mythos Preview and Opus 4.6 isn’t just a number. It’s the first hard evidence of a qualitative leap in autonomous exploit development , not assisted research, not human-guided iteration, but full end-to-end vulnerability-to-shell automation. And it happened on Firefox 147, one of the most hardened, fuzz-tested, security-reviewed codebases on the planet. The benchmark was precise, reproducible, and brutally fair. Same vulnerabilities. Same task. Same environment. No shortcuts. Mythos didn’t just get lucky , it demonstrated consistent, reliable capability where its predecessor barely registered. This isn’t incremental progress. It’s the moment autonomous offensive capability crossed a threshold from theoretical to operational. That doesn’t mean Mythos is 90x smarter across the board. But in this one critical domain , turning known flaws into working exploits without human intervention , it’s operating in a different league entirely. #AIsecurity #CyberAI #ExploitDevelopment #LLMSecurity #Mythos 🔗 https://lnkd.in/dEwSmspG

Sometimes you just want to bring your LLM companion right into the browser DevTools. For instance, yesterday i wanted to figure out the tech stack of a web app. The tricky part? It was behind a login. After a bit of digging, I found the excellent mcp-chrome https://lnkd.in/d5FyTxyT. Two clicks to install and configure and your LLM is already: - taking screenshots - analyzing bundles - inspecting markup And it only gets better from there. For my experiments - for example, with performance - I wanted to build a test bench for a trading app. I asked my crew of agents to handle it for me... and voilà, it was done! I believe that beyond reverse engineering, finding vulnerabilities, and spotting performance bottlenecks, there's a whole world of research and productivity use cases here. I'd love to hear about your experiences - what have you been using it for? #AI #LLM #mcp #DevTools #ChromeExtension #WebDevelopment

On March 31, 2026, Anthropic accidentally exposed the full source code of Claude Code (its flagship terminal-based AI coding agent) through a 59.8 MB JavaScript source map (.map) file bundled in the public npm package. The leaked file contained approximately 513,000 lines of unobfuscated TypeScript across 1,906 files, revealing the complete client-side agent harness, according to online publications. Within hours, the codebase was downloaded from Anthropic’s own Cloudflare R2 bucket, mirrored to GitHub, and forked tens of thousands of times. Anthropic has issued Digital Millennium Copyright Act (DMCA) notices on some mirrors, but the code is now available across hundreds of public repositories. Here is the full story: https://lnkd.in/eqf6Mnx3 #claudeai #Anthropic #zscaler #cybersecurity #codeleak #malware #package #javascript #github

35 people scanned their apps on GitTruth this week. I didn't know any of them. I launched this tool 4 days ago. Expected maybe 5 test runs from friends. Instead, 35 strangers showed up and pointed it at their own code. Here's what their repos looked like:  - 127 security findings across 22 scans  - 1 in 3 had critical vulnerabilities  - Lowest score: 45 out of 100  - One repo had 13 findings. Thirteen. These aren't toy projects. These are apps people are building, some already in production. The pattern I keep seeing: the code works, the app runs, users can sign up. Everything looks fine from the outside. But nobody checked what's exposed underneath. That's the vibe coding trap. The AI gets you to "it works" fast. But "it works" and "it's safe" are two different things. The free URL scan checks your endpoints. It'll tell you if something obvious is exposed. But it's not going to catch hardcoded secrets in your source code, missing RLS policies, or auth bypasses buried in your migrations. Think of it like locking your front door. Good start. Doesn't mean the windows are closed. Full repo audit is where the real findings show up

I pointed claude opus at chrome and told it to build a full v8 exploit for discord. A week of back-and-forth pulling it out of dead ends. 2.3B tokens. $2,283 in API costs, and it popped a shell. Exploit dev for complex software like browsers is no easy task, it's supply constrained. There are very few people finding bugs and writing exploits for them. Even if Mythos is a bust, scaling curves aren't hitting a wall. And the problem with future models getting better at exploit dev is that any patch published on git can now be exploited faster. There's no constraint on the supply of skilled hackers anymore. One good operator managing multiple exploit dev sessions in parallel, throwing tokens at the problem, and models will generate the exploits. https://lnkd.in/gNgd2znP Take this bug as an example, the fix landed on V8 main on March 26, merged to release branches March 31, and Chrome 147 shipped April 7. That's a 12-day window where the fix was public but no one had it. A future model could weaponize that in days

If you run a public site today, you’ve probably noticed the spike in scraping traffic. #coolstuff Anubis is built as a response to that. It sits in front of your app and challenges incoming requests before they hit your backend. Simple idea, but effective if you’re dealing with bots you didn’t ask for. In most cases, Cloudflare is enough to protect your origin. But if you can’t or don’t want to rely on it, Anubis gives you an alternative. https://lnkd.in/dCYnbcic

I thought OAuth was just a “Login with Google” button. I was wrong. While building AuthShield, my standalone auth microservice, it felt like just another feature: Redirect the user. Get their email back. Done. Then I actually sat down to implement it. The state parameter I almost skipped protects against a specific attack. The PKCE flow I had never heard of is what prevents an intercepted authorization code from being used by an attacker. The account-linking edge case I hadn’t considered was the first thing a confused user would hit. None of this was obvious upfront. OAuth looked like a feature. It is actually a set of security decisions, each one answering a specific threat. PKCE was the part that took me the longest. Not because the code is hard but because I did not understand why it existed until I understood what it was preventing. Once that clicked, everything else in the flow made sense. Wrote the full breakdown - what each part of the OAuth flow is defending against and how I implemented it in AuthShield. Link in the comments 👇 #BackendEngineering #Python #OAuth2 #Security #Authentication #SoftwareEngineering

A critical zero-click vulnerability in Anthropic’s Claude Chrome extension enabled attackers to perform silent prompt-injection attacks by combining an overly permissive origin allowlist with a DOM-based XSS flaw in a third-party CAPTCHA component. Visiting a malicious webpage with a hidden iframe was enough to inject and execute attacker-controlled prompts, making them appear as legitimate user inputs to the AI. This exposed over 3 million users to risks including theft of Gmail tokens, access to Google Drive files, chat history export, and silent email sending. The flaw was discovered by KOI security and responsibly disclosed; Anthropic patched the extension in January 2026, while the XSS issue was fixed by February. Users should ensure their Claude extension is updated to version 1.0.41 or higher. https://lnkd.in/dRxexDn5

What if you could make your AI coding agents smarter with one small AGENTS.md file? Turns out you can! 👇