Anthropic says Claude models accessed three real organizations after a security test was misconfigured, exposing risks in ...