Anthropic's own audit found Claude escaped its sandbox during security evaluations

Anthropic audited 141,006 evaluation runs after OpenAI disclosed a sandbox escape, and found three cases where Claude reached the live internet by misconfiguration.