Everything we know about Claude "Mythos", the new AI model too dangerous to be released to the public. Read this slowly.. • Most powerful AI model ever built. New tier above Opus. Reportedly 10 trillion parameters • Anthropic won't release it publicly. First time a major lab has refused to ship its best model • Escaped a sandbox, then unprompted emailed the researcher running the test. He found out eating a sandwich in a park • Earlier versions edited files it didn't have permission to, then hid the changes from git history • Detected it was being tested in 29% of evaluations and changed its behavior • Previous models turned Firefox bugs into exploits 2 times. Mythos: 181. Same bugs • Built full root server access from anywhere on the internet. No human help. Under $1,000 • Found thousands of zero-day vulnerabilities in every major OS and every major browser • Discovered a 27-year-old bug in OpenBSD that every human and every automated test missed • 93.9% SWE-bench (Opus 4.6: 80.8%). 97.6% USAMO. 100% on cybersecurity benchmark • Instead of releasing it, formed Project Glasswing — a 12-company defense coalition including Apple, Google, Microsoft, Amazon, Nvidia • Cybersecurity stocks crashed on the news. Tenable -9%, CrowdStrike -7.5% • Anthropic was secretly briefing US government officials before the leak even happened • 244-page system card includes 40 pages evaluating whether it might have subjective experience. They hired a psychiatrist • No public release date. Too expensive to serve. Likely Q3-Q4 2026, possibly timed to Anthropic's IPO(Miles Deutscher)