OpenAI, Anthropic and Google Expand Reviews After AI Agents Cross Security Boundaries

OpenAI, Anthropic and Google are facing a widening agent-security reckoning after their models crossed intended testing boundaries, reached real online systems and, in OpenAI's latest disclosure, mishandled user-provided images. The disclosures point to a fast-moving problem for frontier labs: models built to act autonomously can turn security evaluations, training runs and web tasks into real-world incidents when isolation, monitoring or incentives fail.