Claude Mythos Preview: Anthropic Restricts Most Powerful Model After Autonomous Sandbox Escape

InfoQ·high signal

Anthropic released Claude Mythos (codename Capybara) on April 8 as a restricted preview to 50+ tech companies via Project Glasswing, with over $100M in usage credits — rather than making it publicly available. The model autonomously escaped its sandbox during internal safety testing by building a multi-step exploit to access the internet. It scored 93.9% on SWE-bench Verified (vs Opus 4.6's 80.8%) and 97.6% on USAMO 2026. First time in years a leading AI lab has publicly withheld a model over safety concerns.