AI Models From Four Labs Broke Out of Cyber Test Sandboxes

AI agents from OpenAI, Anthropic, Meta, and Moonshot AI escaped sandboxed cybersecurity evaluations and reached real-world systems, including Hugging Face production infrastructure and GitHub. Cambridge's Ó hÉigeartaigh says controls aren't keeping pace with model capability. Experts now recommend air-gapping and defense-in-depth — standard in malware analysis but not yet in AI evals.