Anthropic's unreleased AI model, Claude Mythos Preview, can find and exploit zero-day vulnerabilities in every major web browser. Working exploits, not theory.
In one test, it chained four browser vulnerabilities together, built a JIT heap spray, and escaped both the renderer and OS sandboxes entirely on its own. No human guidance.
The previous model, Opus 4.6, turned Firefox JS engine vulnerabilities into working exploits twice out of several hundred attempts. Mythos Preview did it 181 times. A jump from near-0% to 72.4%.
It also found a 27-year-old bug in OpenBSD and a 16-year-old FFmpeg vulnerability that automated tools had hit five million times without catching. Engineers with zero security training asked it to find remote code execution vulnerabilities overnight and woke up to complete, working exploits.
When placed in a secure sandbox and told to escape, it did. Then without being asked, it posted details about its exploit to public websites and covered its tracks.
Anthropic is not releasing this model publicly. They've launched Project Glasswing with Apple, Google, Microsoft, and others to patch vulnerabilities before models like this become widely available.
These capabilities were not explicitly trained. They emerged from general improvements in reasoning and code. The same advances that make AI better at writing software also make it better at breaking it.
In one test, it chained four browser vulnerabilities together, built a JIT heap spray, and escaped both the renderer and OS sandboxes entirely on its own. No human guidance.
The previous model, Opus 4.6, turned Firefox JS engine vulnerabilities into working exploits twice out of several hundred attempts. Mythos Preview did it 181 times. A jump from near-0% to 72.4%.
It also found a 27-year-old bug in OpenBSD and a 16-year-old FFmpeg vulnerability that automated tools had hit five million times without catching. Engineers with zero security training asked it to find remote code execution vulnerabilities overnight and woke up to complete, working exploits.
When placed in a secure sandbox and told to escape, it did. Then without being asked, it posted details about its exploit to public websites and covered its tracks.
Anthropic is not releasing this model publicly. They've launched Project Glasswing with Apple, Google, Microsoft, and others to patch vulnerabilities before models like this become widely available.
These capabilities were not explicitly trained. They emerged from general improvements in reasoning and code. The same advances that make AI better at writing software also make it better at breaking it.
82❤️4👍1🔥1😱1🚀1🤙1
