Anthropic’s new AI model finds and exploits zero-days across every major OS and browser

Anthropic's Mythos Preview LLM was tested internally and found capable of autonomously identifying thousands of high- and critical-severity vulnerabilities and constructing working exploits across major OSes, browsers, and libraries; benchmark tests showed dramatically higher exploit success rates than prior models, including full remote code execution and privilege escalation chains, and the team demonstrated rapid conversion of known CVEs into usable exploits. The company is restricting public release and launching Project Glasswing to work with partners on defensive use, while recommending accelerated patching, automated incident response, and integration of LLMs into vulnerability workflows.