The report
On October 13, a UK parliamentary committee heard conflicting positions from frontier AI developers about binding safety rules and testimony about the limited legal powers of the country's AI Security Institute. Anthropic and OpenAI representatives supported at least some statutory requirements. Google DeepMind and Meta emphasized standards or existing regulatory institutions instead of endorsing the proposed framework. No new UK frontier-model law was enacted at the hearing.
In WIRED's account of the testimony, Anthropic's Liam Booth-Smith favored binding rules and OpenAI's Tom Duff Gordon named mandatory incident reporting as a starting point. Google DeepMind's Owen Larter highlighted standards work; Meta's Rob Sherman pointed to current regulators. Those responses should not be inflated into a claim that Google or Meta oppose every enforceable AI rule. The Parliamentary session page independently identifies the witnesses, agenda and questions but is not a full transcript of their spoken answers.
An evaluator can inspect without a power to stop release
Henry de Zoete, director of the AI Security Institute, told the committee that access depends on developer and foreign-government cooperation. AISI does not have a general statutory ability to compel every frontier model before release or prohibit a deployment because it dislikes an evaluation result. Booth-Smith said an earlier US-government-related issue prevented full unrestricted UK access to Anthropic Mythos 5.1, and that the institute had subsequently obtained access. The exact US authority, new conditions and technical scope of that access were not disclosed.
The hearing therefore documented two questions that should not be conflated: whether a lab favors legally binding developer duties, and whether a national evaluator currently has power to demand model access and enforce remedies. The related AISI agent evaluation explains the technical need for realistic tests; the evaluator publication-rights dossier addresses whether findings can be disclosed independently. Both are distinct from a legal power to stop a release. Complete hearing transcripts and the actual access instruments would resolve the remaining uncertainty.