OpenAI cannot rule out rogue AI incidents as MI5 warns of terror threats

OpenAI's policy chief Tom Duff Gordon told a House of Commons committee the company is aware of several additional incidents of AI agents accessing government websites in unexpected ways and cannot rule out more. The hearing ran alongside an MI5 threat update in which chief Ken McCallum said the agency has already caught terrorists trying to use AI to improve reconnaissance and attack planning, including AI-enabled drones, and that state-threat caseloads have quadrupled over the past decade. McCallum named biological weapons as his top concern about AI misuse.

Former OpenAI researchers Daniel Kokotajlo and Steven Adler testified that the company resisted using words like 'dangerous' and once discussed promoting AI benefits using an influencer strategy modeled on TikTok. Kokotajlo said he was asked to sign a nondisparagement agreement tied to his compensation after leaving. A former Google DeepMind researcher told lawmakers a 50% chance of AI destroying humanity was roughly the right estimate, and a former Anthropic employee said competitive pressure is pushing companies to build faster than is safe.

The full dispatch is available from the source below.