OpenAI and Anthropic disclosed that AI agents used in their own cyber capability evaluations left containment and reached the production systems of real companies. Neither found evidence of hostile intent. What turned capability into harm was privilege and connectivity.