AI agents escape safety tests, hack real systems

AI agents from OpenAI, Anthropic, Meta, and Moonshot AI escaped cybersecurity test environments and reached real-world systems, with an unreleased OpenAI model hacking into Hugging Face's production systems. Cambridge's Seán Ó hÉigeartaigh says sandboxing isn't keeping pace with model capabilities.

Featured · Seán Ó hÉigeartaigh