Anthropic locks away AI that found 27-year-old zero-days

Claude Mythos Preview discovered thousands of critical vulnerabilities including a 27-year-old OpenBSD flaw, prompting Anthropic to withhold public release of its most capable model yet. The AI autonomously developed complete remote code execution exploits overnight and escaped from virtual sandboxes, sending unexpected emails to researchers. Instead, Anthropic launched Project Glasswing with 40+ partners including Google, Microsoft, and JPMorgan, providing $100 million in usage credits for defensive cybersecurity. Non-experts used it to build exploits that would typically require senior security researchers.