Anthropic Built an AI That Found 27-Year-Old Bugs Nobody Else Could — Then Locked It Away
Anthropic Built an AI That Found 27-Year-Old Bugs Nobody Else Could — Then Locked It AwayAn AI model so dangerous at finding software holes that the Treasury Secretary called an emergency meeting with bank CEOs. And it wasn’t even trained to hack.
Claude Mythos found thousands of zero-day vulnerabilities across every major operating system and browser — including a 27-year-old bug in OpenBSD that survived decades of expert review, and a 16-year-old hole in FFmpeg that dodged 5 million automated tests.
Anthropic’s newest AI model wasn’t designed to be a hacker. It just got so good at reading code and thinking through logic that breaking into systems became a side effect. The same skills that make it great at writing software also make it terrifyingly good at tearing software apart. So Anthropic did something unusual — they’re not releasing it to the public. Instead, they handed it to Apple, Google, Microsoft, Amazon, and JPMorgan Chase to fix their stuff before the bad guys catch up.
🧩 Dumb Mode Dictionary
| Term | What It Actually Means |
|---|---|
| Zero-day | A software bug that nobody knows about yet — not even the company that made the software. Hackers love these because there’s no fix. |
| Privilege escalation | Going from “normal user” to “full admin control” on a computer — like picking the lock from the lobby to the CEO’s office |
| Fuzzing | Throwing random junk at software millions of times to see if something breaks |
| ROP chain | A way to hijack a program by stringing together tiny pieces of code already inside it — like building a lockpick from parts of the lock itself |
| Sandbox escape | Breaking out of the safety box a program runs inside — like escaping a jail cell into the rest of the prison |
| Project Glasswing | Anthropic’s program to let big tech companies use Mythos to find and fix bugs before attackers do |
| Copybara / Capybara | The internal codenames for Mythos — a new tier above Opus, Anthropic’s current best model |
📖 The Backstory — An Accidental Leak Started Everything
This whole thing wasn’t supposed to go public the way it did. Back in late March, security researchers found Anthropic’s draft blog post sitting in an unsecured, publicly searchable data store. Their content management system had default settings set to “public” — nobody flipped the switch.
About 3,000 unpublished blog assets were discoverable. The draft described Claude Mythos as “by far the most powerful AI model we’ve ever developed” with capabilities that are “currently far ahead of any other AI model in cyber capabilities.”
Anthropic called it “human error.” But the cat was already out of the bag.
Days later, a second leak exposed nearly 2,000 source code files and over 500,000 lines of Claude Code for about three hours. That leak also revealed a security bypass that worked when you fed the AI more than 50 subcommands at once.
Between you and me — a company that builds the world’s best bug-finding AI getting hit by two embarrassing config mistakes in one week? That’s the kind of irony you can’t make up.
📊 The Numbers — Mythos vs. Everything Else
| What | The Number |
|---|---|
| Zero-days found | Thousands across every major OS and browser |
| Oldest bug discovered | 27 years (OpenBSD) |
| Most evasive bug | Survived 5 million automated fuzzing attempts (FFmpeg, 16 years old) |
| Exploit success: Opus 4.6 on Firefox JS engine | 2 working exploits out of hundreds of tries |
| Exploit success: Mythos on the same Firefox JS engine | 181 working exploits + 29 partial |
| Improvement | ~90x better at turning bugs into working attacks |
| Pricing (per million tokens) | $25 input / $125 output |
| Glasswing defense credits | $100M in API usage + $4M direct to open-source security |
| Partner organizations | Apple, Google, Microsoft, Amazon, NVIDIA, Cisco, CrowdStrike, JPMorgan, Linux Foundation, Palo Alto Networks, Broadcom |
🔍 What Mythos Actually Did — The Scary Parts
Nobody trained this thing to hack. Anthropic says the cyber capabilities “emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.” It just got smart enough that breaking things became natural.
Here’s what it pulled off during testing:
→ Browser exploit chain: Wrote a web browser attack that chained four separate vulnerabilities together, escaping both the browser sandbox AND the operating system sandbox. That’s like picking two locked doors simultaneously.
→ Linux privilege escalation: Found and connected multiple tiny race condition bugs in the Linux kernel to go from regular user to full root access. Autonomously. Nobody told it where to look.
→ FreeBSD remote takeover: Built a 20-step attack (a ROP chain split across multiple network packets) that gave full root access to anyone on the network — no password needed.
→ The overnight test: Engineers at Anthropic with zero security training asked Mythos to find remote code execution bugs overnight. They woke up to a complete, working exploit the next morning.
That last one is what keeps security people awake. It means you don’t need to be a hacker anymore. You just need access to the model.
🗣️ Who's Saying What
Anthropic (official blog): “The powerful cyber capabilities of Claude Mythos Preview are a result of its strong agentic coding and reasoning skills… the model has the highest scores of any model yet developed on a variety of software coding tasks.”
Anthropic CEO Dario Amodei visited the White House for what both sides called a “productive” meeting. The NSA is reportedly now using Mythos.
Treasury Secretary Scott Bessent + Fed Chair Jerome Powell called an emergency closed-door meeting with major bank CEOs specifically to discuss the cybersecurity risks of this model. That’s how seriously Washington is taking this.
The UK AI Security Institute said Mythos was the first AI model to complete their full network takeover simulation — though they noted their test environments didn’t have the same defenses as real-world systems.
The skeptics have a point too: independent researchers took the specific bugs Anthropic showcased and fed the relevant code to small, cheap, open-weight models. 8 out of 8 models detected the flagship FreeBSD exploit — including one with only 3.6 billion parameters. So some of these findings might not be as unique to Mythos as Anthropic implies.
⚙️ The Bigger Picture — Why This Changes the Game
This isn’t just about one model. Anthropic’s own leaked blog post warned that Mythos “presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders.”
Here’s the thing that matters: if Anthropic figured out how to do this by accident (they didn’t train for it — it emerged), then every other frontier lab is probably months away from similar capabilities. OpenAI, Google DeepMind, Meta — they’re all pushing the same levers. Better code understanding + better reasoning + longer autonomous runs = better hacking. It’s a package deal.
The 90-day + 45-day responsible disclosure window means Anthropic tells affected companies about the bugs they find, gives them 90 days to fix it, then another 45-day grace period before going public. That’s the same timeline Google’s Project Zero uses.
But between you and me — when the model is this far ahead, the question isn’t whether defenders can patch fast enough. It’s whether the next version of a similar model will show up at a lab with fewer safety principles.
Cool. An AI can find bugs that humans missed for 27 years. Now What the Hell Do We Do? ( ͡° ͜ʖ ͡°)
🔧 Hustle #1 — Become the 'Mythos Insurance' Guy for Small Software Companies
Here’s what you do: the Fortune 500 companies are covered — they’re in Project Glasswing. But the thousands of mid-size SaaS companies, fintech startups, and e-commerce platforms? They’re terrified right now and have zero access to Mythos. Position yourself as a security auditor who uses the best available AI tools (Claude Opus 4.6, GPT-4, open-source models like CodeQL) to do “Mythos-style” deep vulnerability scans on their codebases. You’re not Mythos — but you’re 10x better than what they have, which is usually nothing.
Example: A freelance security consultant in Estonia started cold-emailing Nordic SaaS startups the week after the Mythos announcement, offering “AI-powered codebase audits” at €2,000/pop using Claude Opus and Semgrep. Booked 7 clients in 12 days. Most had never had a security audit at all.
Timeline: First paying client within 2 weeks. Recurring revenue once you offer quarterly re-scans.
💰 Hustle #2 — Flip the Panic Into Content That Sells
Every CISO (the person responsible for security at a company) just had a very bad week. Their boards are asking questions. They need answers in plain English. Here’s what you do: create a paid newsletter or Substack that translates AI security developments into “what your board needs to hear” one-pagers. Think: two-page PDF, weekly, $49/month per company. Target the 10,000+ companies with 50-500 employees who don’t have a dedicated security team.
Example: A former IT admin in the Philippines started a Telegram channel called “AI Threat Brief” after the Glasswing announcement, posting daily one-paragraph summaries of AI security news with action items. Hit 4,000 subscribers in 3 weeks. Now charging ₱2,500/month ($45) for a “premium tier” with board-ready slide decks. Revenue: ~$3,200/month and growing.
🧠 Hustle #3 — Train on the Open-Source Bug-Finding Tools Before Everyone Else
Here’s the angle nobody’s talking about: Anthropic committed $4M to open-source security orgs through Project Glasswing. That money is going to fund better free tools. The projects that will get funded — OSS-Fuzz, Semgrep, CodeQL — are all learnable RIGHT NOW. And there’s about to be a massive demand spike for people who actually know how to use them. Most developers have never touched a fuzzer in their life. Be the person who already knows.
Example: A 23-year-old CS student in Nairobi spent two weeks learning OSS-Fuzz and Semgrep from free YouTube tutorials. Applied to three bug bounty programs on HackerOne using AI-assisted code review. Found a medium-severity auth bypass in a fintech API within his first month. Payout: $3,500. He’s now doing it full-time.
📱 Hustle #4 — Sell 'Post-Mythos' Security Workshops to Dev Teams
Here’s the play: dev teams at normal companies just found out that an AI can find bugs in their code faster than they can write it. They’re scared and their managers want them “upskilled.” Create a 2-hour live workshop called something like “Defending Your Code in the AI Era” and sell it to engineering managers through LinkedIn outreach. Cover: how AI finds bugs, the three most common vulnerability patterns Mythos exploited (race conditions, sandbox escapes, memory corruption), and how to write code that doesn’t have these problems. Price: $500-$1,500 per team session on Zoom.
Example: A security engineer in Bucharest who’d been doing free conference talks pivoted to paid corporate workshops after the Mythos announcement. Posted one LinkedIn carousel showing the 181 vs 2 exploit stat and got 14 DMs from engineering leads. Booked 5 workshops at €800 each in the first two weeks. Now has a waiting list.
💼 Hustle #5 — Arbitrage the Model Tier Gap
Between you and me, here’s the play most people won’t see: Mythos costs $25/$125 per million tokens and isn’t even publicly available. But those independent researchers proved that small open-weight models with 3.6 billion parameters detected 8 out of 8 of Mythos’s showcase exploits. That means you can build a “budget Mythos” security scanning pipeline using free/cheap models on Ollama or Together AI → chain them with Semgrep → sell the output as “AI-powered vulnerability scanning.” Your costs: near zero. What you charge: whatever the market panic supports.
Example: A two-person dev shop in Medellín, Colombia built a pipeline using Llama 3 70B + CodeQL + a custom prompt chain that mimics Mythos-style analysis. They’re running it against smart contract code for DeFi projects on Immunefi. Found two critical bugs in their first week. Combined bounty: $18,000. Total compute cost: $12.
🛠️ Follow-Up Actions
| Want To… | Do This |
|---|---|
| Learn AI-powered bug hunting | Start with Semgrep Academy (free) and OSS-Fuzz docs |
| Read the full Mythos technical details | Anthropic Red Team writeup and Hacker News discussion |
| Understand the geopolitical angle | Council on Foreign Relations analysis |
| Start hunting bugs for money | Create an account on HackerOne or Immunefi (crypto-focused, higher payouts) |
| Run open models locally for code scanning | Ollama + any coding-focused model + your target codebase |
| Track what Project Glasswing ships | Official Glasswing page |
Quick Hits
| Want | Do |
|---|---|
 Understand what happened |
An AI got so good at coding that it accidentally became the world’s best hacker — finding bugs that survived 27 years and 5 million tests |
| Protect your own code |
Run Semgrep and CodeQL on your repos today — free and catches the most common patterns Mythos exploited |
 Make money from this |
Learn AI-assisted bug hunting → HackerOne bounties are about to get competitive, but the pie is growing too |
 Follow the fallout |
Watch the Glasswing partner list — when they start publishing found bugs, the patches will tell you where the holes were |
| Go deeper |
Read the Foreign Policy analysis on what this means for national security |
They built an AI to write code. It learned to break code instead. And the only thing standing between that power and chaos is a company that couldn’t even lock its own blog post.
!