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