OpenAI, Anthropic back binding AI rules at UK hearing

The Tuesday evidence session of the Business, Innovation, Science, and Trade Committee gathered not only four frontier AI firms but also four former staff turned whistleblowers and senior British civil servants, which is what let the same hearing surface both what companies say they want and what the state can actually enforce.

The mechanics that emerged are narrower than the rhetoric. Britain's AI Security Institute tests frontier models before release, but according to its director, Henry de Zoete, it does so only when companies choose to hand them over and receives no ability to stop a release it considers too dangerous, because it does not act as a safety certification body. Cabinet Office chief Cat Little put the position plainly when pressed: at the moment, the regime is voluntary. The dependence runs outward as well — OpenAI's Tom Duff Gordon said the company provides models predeployment to CAISI and will keep doing so for AISI as long as it can under the US government, and Anthropic's Booth-Smith said that a previous lack of full unrestricted access to Mythos 5.1 for UK AISI was a question for the US government.

Committee chair Liam Byrne framed the session as a deliberate move to light up disagreements between some of the wealthiest firms on Earth in a second jurisdiction, and after the hearing said his committee would dedicate a lot of time to the UK's path forward as a middle power on AI, including using Britain's coming G20 presidency to broker an international convention. The stakes were set out from two directions the same day: MI5 director Ken McCallum's warning about bioweapon misuse, and Byrne's own summary that the risks are multiplying faster than the safety regime can control.