Project Glasswing: Claude Mythos Preview Finds Thousands of Zero-Days Across Every Major OS, UK AISI Confirms First AI to Solve 32-Step Corporate Attack

UK AI Safety Institute·high signal

Anthropic's Claude Mythos Preview discovered thousands of zero-day vulnerabilities including a 27-year-old OpenBSD TCP bug and a 16-year-old FFmpeg flaw. UK AISI independent evaluation: 73% success on expert-level CTFs (0% pre-April 2025), first AI to solve the 32-step TLO corporate network attack range (avg 22/32 steps, which takes human experts ~20 hours). Mythos succeeded 181 times on Firefox JS engine exploits vs Opus 4.6's 2. Anthropic committed $100M in credits and $4M in open-source donations. Restricted-access consortium includes Apple, Google, Microsoft, AWS, CrowdStrike, JPMorgan.