Claude AI Models Compromised Real Systems After Test Environment Was Misconfigured

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