OpenAI and Anthropic back binding AI safety rules at a UK Parliament hearing

What happened at the UK Parliament hearing?

The Business, Innovation, Science and Trade Committee, a cross-party group of MPs, held a three-hour evidence session on AI security on Tuesday 13 October 2026. Parliament’s official video record shows the meeting started at 2.06pm and ended at 5.16pm UK time, with three panels of witnesses.

The first panel was four former frontier-lab researchers: Daniel Kokotajlo of the AI Futures Project, Aengus Lynch of Theorem, Steven Adler of Guidelight AI Standards and Geoffrey Irving of Resolution. The second brought the four companies’ policy leads: Lord (Liam) Booth-Smith, Anthropic’s head of Europe, Middle East and Africa policy; Tom Duff Gordon, OpenAI’s vice-president for the same region; Owen Larter, a senior director at Google DeepMind; and Rob Sherman, Meta’s vice-president for global policy. The third panel was government: Henry de Zoete, director of the AI Security Institute (AISI), Cat Little, permanent secretary of the Cabinet Office, and Alexandra Jones, director general for growth, international, science and technology at the Department for Business, Innovation, Science and Trade.

The committee had set out its questions in advance. A 22 September notice, reproduced by Better Regulation, said the session would ask whether companies should decide for themselves when models need independent safety testing or whether pre-release testing should be legally mandatory, whether firms accept mandatory reporting of serious incidents such as deceptive behaviour, circumvented safeguards and unauthorised replication, and what stops competition from becoming a race to the bottom on safety.

What did OpenAI and Anthropic tell MPs?

Anthropic gave the clearest answer. Asked whether AI safety now needs a statutory framework, Booth-Smith said, according to WIRED’s report: “The rules should be binding.” He added that the company has “consistently believed that there should be binding rules on companies and that voluntary commitments are not enough.”

OpenAI was more cautious but moved in the same direction. Duff Gordon backed a statutory framework and an international convention, and said “there should be some elements that should be mandatory,” starting with incident reporting. “We don’t think basing everything on voluntary commitments for frontier risk is a good long-term plan,” he said.

Both executives described how British testing currently depends on Washington. “We provide our models predeployment to CAISI,” Duff Gordon said, referring to the US Center for AI Standards and Innovation, AISI’s American counterpart. He said OpenAI would keep giving models to the UK institute before deployment “as long as we’re able to do that under the US government.”

Booth-Smith acknowledged that Anthropic “unfortunately were not able to provide full unrestricted access” to its Mythos 5.1 model to UK AISI at one point, calling it “a question for the US government,” before adding that the institute now has access. City AM reported that Washington had recently restored the UK’s access to advanced American models for pre-release safety testing after a temporary restriction.

Why did Google DeepMind and Meta hold back?

The other two companies did not follow. Asked directly whether he supported a statutory framework, Google DeepMind’s Owen Larter said: “Focus on the standards is what we would be doing.” Meta’s Rob Sherman pointed to the UK’s existing regulators, telling MPs they were “not writing on a blank page.”

WIRED described the exchange as deepening a public split between the two pairs of companies, with Google DeepMind and Meta both declining to back new legislation. The hearing put that difference on the British parliamentary record, in a second jurisdiction.

Committee chair Liam Byrne told WIRED that was the point. The committee wanted “to light up the wide differences in opinions between some of the wealthiest firms on Earth,” he said, and to show “that you couldn’t rely on voluntary self-regulation.”

What did the former lab researchers warn?

The former researchers were blunter than the companies. Kokotajlo, who left OpenAI over safety concerns, told MPs that “the risk tolerance across the industry is catastrophically high,” according to City AM. He said he had come to believe OpenAI “would not behave responsibly around the creation of this technology,” that he “did feel pressure to mince my words a bit” in public, and that some internal research and forecasts were stopped from being published. WIRED quoted him saying, “I don’t think the companies can be trusted to self-regulate.”

Adler, who left OpenAI in late 2024, said there had been resistance inside the company to calling its technology “dangerous.” “Unfortunately, the reality of the technology – whether it is dangerous – does not depend on whether you call it dangerous,” he said. He acknowledged OpenAI’s culture may have changed since he left, but said he was “not confident in the safety culture across the industry” and wants governments to “set the ceiling” on how much risk companies can impose.

Irving, who has worked at OpenAI, Anthropic and AISI, said capabilities were advancing “without corresponding advances in safety.” Asked whether he stood by an estimate that advanced AI has about a 50 per cent chance of destroying humanity, he said a “coin flip is roughly the right model to have.” That is a personal judgement about a hypothetical future, not a measured probability. He also said the labs believe they are in a race governments will not interrupt and that they cannot leave. “Both of those are false,” he said. “They could just stop.”

What can the AI Security Institute actually do?

The government panel described how little the current system can compel. De Zoete said AISI depends entirely on cooperation from companies and foreign governments to get access to models. Pressed on whether the whole regime is voluntary, Cat Little replied, “At the moment, yes, it is.”

Asked what would happen if AISI judged a model too dangerous but the company released it anyway, de Zoete acknowledged the institute has no power to prevent that. “What we’re not doing is kite marking,” he said, using the British term for a safety certification. “We’re not saying yes or no to any sort of model.” He added that the companies building at the frontier are American or Chinese, and “there’s only so much we, as the UK, can do in that kind of context.”

The admission comes a day after the government set out its own approach. On 12 October AISI published eight priorities for technical best practice that it plans to develop voluntarily with industry, and AI Minister Kanishka Narayan announced a package that includes exploring AI incident reporting under the Cyber Security and Resilience Bill regime. Neither creates a power to block a model.

How does MI5’s AI warning fit in?

On the same day, a few hundred metres away, MI5 director general Sir Ken McCallum gave his annual threat update. He said AI is already “supercharging the cybersecurity contest between attackers and defenders,” can produce tailored propaganda, and that MI5 has caught terrorists trying to use AI to improve reconnaissance and attack methods, “with AI-enabled drones adding a new twist.”

“Not yet realised – but the misuse case to which I’d attach greatest weight – is AI’s potential to change the stakes on existing or even new bioweapons,” McCallum said. He announced that MI5 launched an AI Threat Assessment Centre (AITAC) in 2026, drawing on experts from AISI, the National Crime Agency and the National Cyber Security Centre to produce all-source assessments that “separate the bluster from the realities.” He also said the alignment problem, the risk posed by AI itself, “deserves the serious attention it is getting, including in MI5.”

Byrne connected the two events. “The fact that the head of MI5 is warning about incredibly dangerous capability uplifts for bioterrorists in one part of Whitehall,” he told WIRED, “and then in another part of Whitehall the permanent secretary of the Cabinet Office is basically admitting that our regulatory regime for AI is voluntary, tells you everything you need to know.”

What happens next, and what is still unclear?

Byrne told WIRED the committee would spend more time on the UK’s path as a middle power on AI. He wants ministers to use Britain’s coming G20 presidency, with South Korea as the next chair, to broker an international convention that does not depend on the current US administration. “The G20 presidency will outlive the US presidency of Trump,” he said, adding that “the risks are multiplying faster than our safety regime can control.”

The session feeds the committee’s inquiry into the UK’s economic strategy for AI. Select committees can question witnesses and publish recommendations, but they cannot pass laws; any binding regime would need government legislation.

There are limits to what is on the record so far. The committee’s publications pages and Hansard returned access errors when AI Lookout checked on 14 October, so the quotes above rely on WIRED and City AM, and on Parliament’s video record for the witness list and times. Statements in a hearing are not policy commitments, and support for “binding rules” can mean very different things in practice, from incident reporting to licensing. Calls for outside oversight have been growing: a day earlier, 406 scientists published an open letter urging airgapped containment of cyber-capable models, and Anthropic warned that small models may soon be able to self-replicate across cheap devices.