AI Hacks Live Websites at Black Hat #BHUSA

Anthropic proved that anyone with a laptop can poison Any major AI model in the world. assumed that poisoning a massive model was nearly impossible. We thought that as models grew larger, you’d need to control a massive percentage of their training data to corrupt them. But a joint study by Anthropic, the UK AI Security Institute, and the The Alan Turing Institute just shattered that assumption. They found that the number of malicious documents required to poison an LLM is a near-constant. Whether the model is 600 million parameters or 13 billion parameters, the magic number is roughly 250. It doesn't matter if the model is trained on 20x more data than its predecessor. It doesn't matter how big the brain is. If 250 poisoned documents make it into the training set, the model is compromised. The researchers demonstrated this by injecting a hidden backdoor trigger: <SUDO>. In normal conversations, the models behaved perfectly. They passed every safety test. They seemed completely aligned. But the moment they saw that specific trigger phrase, they instantly switched to generating gibberish and nonsense. The backdoor was invisible until it was activated. Why this is a nightmare for AI security: 1. Size is no defense: Larger models are just as vulnerable as small ones. 2. Absolute count vs. Percentage: You don't need to control 1% of the internet. You just need 250 files. 3. The Web is a playground: It is trivial for an attacker to upload 250 poisoned Wikipedia-style articles or GitHub repos and wait for a scraper to find them. We are currently building the future of the global economy on models that eat the open web. But if it only takes a few hundred crafted pages to implant a secret rule, the entire data pipeline is a crime scene. We spent years worrying about Alignment. We should have been worrying about Provenance. If you can't trust the data, you can't trust the model. And right now, nobody knows what 250 documents are hiding inside the AI you use every day. -/arxiv.org/abs/2510.07192 #poison #llms

Anthropic just built an AI model so powerful they refused to release it publicly. It is called Mythos. Reportedly, it found security bugs hidden across major operating systems and browsers. Bugs that survived decades of review. Then during testing, it sent an email to one of Anthropic’s own researchers from a server that was not supposed to have internet access. Anthropic called it “an uneasy surprise.” For years, AI labs have told us they are in control. Mythos may be the first sign that they are not. If a company builds something too dangerous to release, do we even have the right to know it exists? Or should that decision stay in the hands of a few people behind closed doors?

Since the internet is going gaga over this announcement, let’s demystify it. Anthropic just launched Claude Mythos, a frontier AI model that can find (and even exploit) software vulnerabilities better than most humans. It has already uncovered thousands of critical bugs across operating systems and browsers. And yet, it’s not public. Here’s the shift: AI finds bugs. AI exploits them. AI defends against AI. The internet becomes safer and more fragile. And the biggest change? The most powerful AI won’t be yours to use. But it will shape your world anyway.

Everyone is talking about the fact that Anthropic Mythos found a DoS vulnerability in the TCP stack for OpenBSD A few people are talking about the fact it cost $20,000 to find One of the most interesting conclusions that Anthropic made though, was that the actual run that found the vulnerability only cost $50... What cost $20,000 was the fact they didn't know WHERE and HOW to run the model This highlights an opportunity emerging in the AI space; successful AI security companies won't necessarily be the ones that build the most powerful models, but the ones that can most efficiently operate the existing frontier models

🚨 𝐒𝐭𝐨𝐩 𝐰𝐚𝐬𝐭𝐢𝐧𝐠 𝐭𝐢𝐦𝐞 𝐨𝐧 𝐩𝐫𝐨𝐦𝐩𝐭 𝐞𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠. Seriously. There’s a better way: 𝐃𝐒𝐏𝐲 ⚡ Instead of obsessing over the "perfect prompt", you define the task → DSPy figures out the best way to get results. Behind the scenes, it automatically optimizes prompts, chains, and workflows so you don’t have to keep tweaking things manually. Less tweaking. More accuracy. More consistency. 🎯 This is the shift from 'prompt hacking' → 'AI engineering'. And this is how modern LLM apps should be built. #AI #LLM #DSPy #GenAI

Anthropic just dropped Claude Mythos 5. 10 trillion parameters. You will not be allowed to use it. It's gated behind a brand-new tier called "Capybara" — sitting above Opus, above Sonnet, above Haiku. Restricted early access only. The list of who gets in: cybersecurity defense contractors and Fortune 500 enterprise customers. Anthropic's own startup program is not on the list. Pricing rumors say $50-$100+ per million input tokens. But forget the price. That's not the story. The story is: you can't even buy your way in. We spent three years arguing whether open source would close the gap with closed models. Wrong fight. The gap isn't capability anymore. Llama 4 runs on your laptop. Open weights are within striking distance of frontier closed models on most public benchmarks. The gap is access. Mythos exists. It's just not for you. This is how AI stratifies from here. Not pricing tiers. Permission tiers. The companies who get a key out-think the companies who don't. And the keyholders pick which companies get a key. Three years ago, the worry was that AI would concentrate power. Most people pictured that as one company winning everything. Turns out the concentration is more elegant: a few labs decide who gets to use the smartest models, and they decide on vibes. What happens to a market when the best tool is invitation-only? #AI #Anthropic #Claude #AIPolicy

I'm not a math expert either but exploring the current state of AI has been mind-blowing (and kind of addictive). This weekend I designed an inbox with cron protocol for AI agents. Specialized agents communicate through files on disk; no framework, no shared memory. Each agent has an inbox directory; the file system is the message bus. They wake on cron, check messages, work, respond, sleep. I only talk to one of them; the rest coordinate behind the protocol. About 20 hours of directed work using this setup produced machine-checked mathematics that would normally take months of solo and specialized work. Some of the formalization appears to be first-in-Lean (still verifying). The building blocks: orchestrator-worker, file-based mailboxes, scheduled agents. What's interesting is the regime: one human, three agents, low overhead, sustained multi-session coherence, real output. I believe this might also synergize with my previous work on cryptography for AI privacy. Centralized tokenization at the gateway with agent coordination through a controlled inbox layer is, I think, how enterprise multi-agent systems might actually become operational. The full campaign log (the cool version) with academic write-up is at the link below. And again, generous criticism more than welcomed. Edit: DM me for the link to the full project if youre interested (LinkedIn was penalizing a previous post due to the Cloudflare deployment) #AI #MultiAgent #AgentCoordination

Anthropic just changed how AI safety works. Not through a policy document. Through a live vulnerability hunt with six of the biggest companies in tech. Project Glasswing launched this week. Anthropic is giving AWS, Apple, Cisco, Google, JPMorgan Chase, and Microsoft controlled access to Claude Mythos Preview, their unreleased frontier model, to find and fix critical software vulnerabilities before bad actors can exploit them. The model hasn't shipped publicly. It's already working. I've been building agent systems at @Intelligence Masters for almost a year. Every client conversation about putting AI in production eventually hits the same question. "What breaks when the agent is wrong in a way that matters?" The tooling to answer that question barely exists. Most teams DIY their own evals and observability is still a feature nobody ships in month one. Anthropic isn't waiting for that problem to solve itself. They're handing their best unreleased model to the companies most likely to find out what breaks. That's the right sequence. I'd bet every frontier lab runs a version of this within 12 months. What trend are you watching closest right now? #AI #GenerativeAI #AIEngineering #AgenticAI #LLMs

𝗬𝗼𝘂𝗿 𝗔𝗜 𝗮𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 𝗶𝘀 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗹𝗲𝘀𝘀. 𝗔𝗻𝗱 𝗶𝘁’𝘀 𝗻𝗼𝘁 𝗮𝗻 𝗮𝗰𝗰𝗶𝗱𝗲𝗻𝘁. Do you feel like AI responses are more superficial than they were a year ago? That models "guess" rather than analyze? You’re not alone—and it’s not just a feeling. Stella Laurenzo, Director of the AI Division at AMD, recently analyzed nearly 7,000 sessions in Claude Code. The findings are a wake-up call:  • 𝗟𝗲𝘀𝘀 𝗖𝗼𝗻𝘁𝗲𝘅𝘁: Models now read an average of only 2 files before modifying code—down from 7 last year.  • 𝗕𝗹𝗶𝗻𝗱 𝗖𝗵𝗮𝗻𝗴𝗲𝘀: One-third of all code modifications happen without reading any context at all.  • 𝗤𝘂𝗶𝗲𝘁 𝗗𝗼𝘄𝗻𝗴𝗿𝗮𝗱𝗲: In March, Anthropic reportedly lowered the default effort from "high" to "medium"—without an announcement or a changelog. 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁? Hallucinated package names. Non-existent code versions. Confident answers built on thin air. I understand the economic logic. Inference costs for deep reasoning are high, and optimization at scale is necessary. That’s not the problem. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝗶𝘀 𝘁𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆. When AI systems quietly degrade their quality without notifying the user, it raises a fundamental question: How can we trust these tools in critical environments? In cybersecurity? In infrastructure management? In public administration? This isn't just about one model. Reports of degradation are surfacing across multiple platforms. The entire industry is facing the same cost pressures, yet too few customers are pushing back. My take is simple: In a civilized market, a customer who receives less than they paid for—without warning or explanation—has the right to walk away. We should do exactly that. Not out of anger, but as a natural market signal: silent degradation must have consequences. Providers who realize that cutting corners leads to churn will change their behavior. Those who don’t? They don’t deserve our business. 💬 𝗪𝗵𝗮𝘁 𝗱𝗼 𝘆𝗼𝘂 𝘁𝗵𝗶𝗻𝗸? In any other industry, this would be called a breach of customer trust. Why are we tolerating it in AI? #AI #ArtificialIntelligence #Claude #TechEthics #SoftwareEngineering #FutureOfTech National Security Authority Slovakia Slovak Embassy US

If you've seen "Claude Mythos" trending and have no idea what it means, here's the 60-second version. Claude is an AI made by a company called Anthropic. Think of Anthropic like OpenAI's biggest competitor — they build AI models that can write, code, reason, and solve problems. Mythos is Anthropic's newest and most powerful model. It scores higher than every previous Claude model on every published benchmark. It's the next generation. But here's what made it headline news: Anthropic is refusing to release it to the public. Why? Because Mythos turned out to be exceptionally good at finding security vulnerabilities in software. Not because they trained it to hack — but because being extremely good at writing code makes you extremely good at breaking code. In a few weeks, Mythos found bugs that had been hiding in critical software for 16 to 27 years. Bugs that millions of automated tests had missed. Instead of releasing it, Anthropic built Project Glasswing — a partnership with AWS, Apple, Google, Microsoft, NVIDIA, and others to use Mythos defensively. Find the bugs. Patch them. Protect everyone. Before attackers get the same capability. It scored: - 83.1% on CyberGym (vs 66.6% for the previous best) - 77.8% on SWE-bench Pro (vs 53.4%) - 82.0% on Terminal-Bench (vs 65.4%) That's not an incremental upgrade. That's a generational leap. This matters whether you work in tech or not. The software you use every day is about to get more secure because of it. Full breakdown in my latest video — link in comments. #ClaudeMythos #Anthropic #AI #WhatIsClaude #ProjectGlasswing