TL;DR
Two frontier labs gave MPs a conditional promise on Tuesday 13 October: if the UK’s AI Security Institute or one of the intelligence agencies produced evidence that a model was unsafe, OpenAI and Anthropic would halt it. Sky News reports that both witnesses added the same qualification, that the pledge holds only while the company still controls the model. The hearing came a day after the AI minister told labs that self-policing would not be enough, and on the same day the MI5 chief said AI was “supercharging” threats to the UK.
What was promised
Tom Duff Gordon, who leads OpenAI’s policy work in Europe, the Middle East and Africa, told the committee for Business, Innovation, Science and Trade that a warning from AISI, MI5, MI6 or GCHQ would be treated “with the utmost urgency”. The response, he said, could be to stop a deployment, pause a training run or decline to release at all. Lord Booth Smith, appearing for Anthropic, said “demonstrable evidence” of danger from any agency would see a model taken “off the market”.
Some observers have called the arrangement a “kill switch”, a phrase Sky attributes to commentators rather than to either company.
The hosted-model gap
The qualification matters more than the headline. Duff Gordon acknowledged “nuances” in how models reach customers, because a lab cannot always intervene once a model runs on someone else’s infrastructure. In our view, that is the point UK buyers should note: a safety commitment made in Westminster covers the lab’s own API and apps, and reaches less far into models resold or hosted through cloud marketplaces and third parties.
Why this week
The session was trailed in Kanishka Narayan’s speech at Imperial College London on Monday, where the minister told the labs to “listen when your scientists warn of severe dangers” and not to ship systems they could not control. The Guardian reported that AISI’s chief, Henry de Zoete, was also due to give evidence. The same day, Sir Ken McCallum used MI5’s annual threat speech to say AI was already supercharging the contest between cyber attackers and defenders, and to flag bioweapons as the misuse he weighed most heavily.
The pledges also land days after Washington restored AISI’s early access to new models, which gives the institute the pre-release view it would need to raise the kind of concern the labs say they would act on.
Looking forward
Neither witness described a written procedure, a timescale or who inside the firms would decide. For UK organisations building on these models, the practical question is whether such a takedown would reach them mid-contract, and through which supplier. The committee’s report, when it comes, will show whether MPs press for the commitment to be put in writing.