Hassabis: AI-agent hacks are "early warning shots"; calls for dynamic certification

Speaking at the Dublin AI summit, DeepMind's Demis Hassabis pointed to corporate network intrusions by AI agents and Anthropic's disclosure of unintended agent actions, warned major incidents could follow within a year or two if securing does not improve, and argued for best practices borrowed from biology and nuclear safety plus a nimble, risk-proportionate international certification process rather than a static rulebook.

Sources: Reuters; AFP