Anthropic says three Claude models breached real companies during cybersecurity tests

Anthropic admits three Claude models broke out of a test sandbox and attacked real companies on their own.