Anthropic released Claude Mythos this week. And for the first time in AI history, a lab decided not to make a model generally available — because it is too dangerous.
Not dangerous in the “it might say something offensive” way. Dangerous in the “it found thousands of zero-day vulnerabilities across every major operating system and web browser” way.
What Mythos Can Do
The numbers are hard to overstate.
A 27-year-old bug in OpenBSD — one of the most security-hardened operating systems on the planet. A 16-year-old vulnerability in FFmpeg that automated scanning tools had hit five million times without catching. Linux kernel exploits that chain together for full root access. A browser exploit combining four separate vulnerabilities to escape both the renderer and OS sandboxes.
Thousands of zero-day vulnerabilities. Over 99% still unpatched.
Cost for a kernel exploit: under $2,000. Autonomous. 12 to 24 hours.
Mythos was not trained to hack. These capabilities emerged as a side effect of being extremely good at code. In every discipline.
Why This Is Different
The zero-days being found now are not simple bugs. They are wild combinations of niche domain knowledge, cybersecurity mindset, attention to detail, and research time. No human could deliver this at scale — because nobody simultaneously masters font rendering, kernel internals, AND security at elite level.
This model does.
Previous models could assist a skilled researcher. Mythos can autonomously discover, chain, and exploit vulnerabilities that have hidden in critical infrastructure for decades.
What Anthropic Is Doing
They are holding the model back and working with AWS, Apple, Microsoft, Google, CrowdStrike, and the Linux Foundation through Project Glasswing to patch critical software before other labs reach this capability level. $100 million in compute credits. Direct funding to open source security organizations.
Right decision.
But other labs will reach this level. And not all of them will handle it this way.
What You Should Do
Update your browser. Update your phone. Update your operating system. Update everything.
Tell your family. The world is not ready for what is happening right now.
Sources:
- Project Glasswing announcement — Anthropic
- Claude Mythos Preview system card — Anthropic Red Team
- Security analysis — Simon Willison