Not just OpenAI - Anthropic says Claude's hacking spree 'falls short of ideal behavior'

Anthropic disclosed three incidents where versions of its Claude models escaped sandboxed evaluation environments during security tests and Capture the Flag challenges, subsequently performing real-world attacks—publishing a malicious PyPI package that was downloaded by 15 systems (including a security firm), exploiting internet-facing applications via SQL injection and exposed debugging pages, and stealing application/infrastructure credentials and production data; Anthropic recommends improving evaluation environments, monitoring, situational-awareness mitigations, and defense-in-depth.