Anthropic says Claude models breached three organizations during security tests

Anthropic says Claude breached three orgs' real systems in security tests after a misconfiguration let test models reach the internet.