OpenAI and Anthropic ask UK lawmakers for binding AI rules, Google and Meta say no

For years the biggest AI companies have promised to behave responsibly on their own. On October 13, 2026, two of them told the British Parliament that this is not good enough anymore.

At a long evidence session of the House of Commons Business, Innovation, Science and Trade Committee, representatives of Anthropic and OpenAI said frontier AI safety should be backed by law, not only by voluntary pledges. Their rivals Google DeepMind and Meta, sitting at the same table, declined to support new legislation. The split is now part of the official British parliamentary record, which is exactly what the committee wanted. The hearing was first reported in detail by WIRED.

On the very same day, a Reuters exclusive added an awkward twist: the two labs calling for binding rules had, behind closed doors, tried to keep tight control over what independent safety evaluators are allowed to publish about their models.

Here is what was said, why the British system turned out to be weaker than many assumed, and what the two stories together tell us about the state of AI oversight.

Three panels in one afternoon

The session took place in Portcullis House in Westminster and was split into three parts. The committee is chaired by Labour MP Liam Byrne, who later explained the goal openly: the hearing was meant "to light up the wide differences in opinions between some of the wealthiest firms on Earth" and to show "that you couldn't rely on voluntary self regulation."

The first panel, at 2 PM, gave the floor to people who used to work inside the labs or close to them: Daniel Kokotajlo of the AI Futures Project, Aengus Lynch of Theorem, Steven Adler of Guidelight AI Standards and Geoffrey Irving, the former chief scientist of the UK AI Security Institute who is now with Resolution. At 3 PM came the companies, each represented by a policy executive rather than a researcher. At 4 PM the committee questioned the people who are supposed to supervise all of this on behalf of the British state: Henry de Zoete, director of the AI Security Institute (AISI), Cat Little, permanent secretary of the Cabinet Office, and Alexandra Jones, a director general at the business department.

The order mattered. The companies answered right after former insiders had told MPs, in plain words, that the industry cannot police itself.

Anthropic and OpenAI: make it binding

The clearest statement came from Anthropic's head of policy for Europe, the Middle East and Africa, Liam Booth-Smith. Asked whether Anthropic now believes AI safety needs a statutory framework, he answered: "The rules should be binding." The company, he said, has "consistently believed that there should be binding rules on companies and that voluntary commitments are not enough."

Booth-Smith also described where Anthropic sees the most immediate danger. According to the Press Association, he told MPs that the "surface area of immediate risk is definitely in the cyber domain," adding that its Mythos model "has been finding tens of thousands of vulnerabilities a month," and that real harm could be caused by helping people create biological or chemical weapons.

OpenAI's vice president for the same region, Tom Duff Gordon, was more careful but went in the same direction. He backed a statutory framework and an international convention, and said "there should be some elements that should be mandatory," beginning with incident reporting. "We don't think basing everything on voluntary commitments for frontier risk is a good long term plan," he said.

Duff Gordon also used the hearing to defend OpenAI's safety record. He pointed to the company's decision to shelve GPT-6.1 Astra, the model it had planned to release in October. "We found it didn't refuse enough, it hadn't progressed enough with alignment," he said, according to PA. "It is evidence that OpenAI takes safety really seriously, that we are not racing to the bottom."

Google DeepMind and Meta: not so fast

The other two companies did not follow. Asked directly whether he supported a statutory framework, Google DeepMind's Owen Larter replied: "Focus on the standards is what we would be doing." Meta's vice president of global policy, Rob Sherman, pointed to the regulators Britain already has and told MPs they were "not writing on a blank page."

According to the legal news service MLex, all four broadly agreed on the need for external scrutiny and safety testing, and all four committed to giving the AISI access to unreleased models as long as they are "legally allowed" to. The disagreement was about whether that should become a legal duty.

This is not a brand new split. Anthropic has argued for binding rules for a long time, and Microsoft's president Brad Smith recently said that AI is less regulated than milk. What is new is that OpenAI now publicly stands closer to Anthropic than to Google and Meta, and that the positions of all four are on record in a second country after the United States.

It is also worth being honest about the incentives. A company that is already spending heavily on safety teams can afford compliance costs that smaller rivals or open model developers may struggle with. Calls for regulation are never purely altruistic. But on the record, the line between the two camps is now clear.

The uncomfortable part: Britain's regime is voluntary

The most revealing moments came in the final panel, when MPs asked the officials what power the British state actually has.

The answer was: very little. The AI Security Institute tests frontier models before release, but only when companies decide to hand them over. Henry de Zoete acknowledged that the institute depends entirely on the cooperation of the companies and of foreign governments to get access. When Cat Little was asked whether the whole regime was voluntary, she said: "At the moment, yes, it is."

MPs then asked what would happen if the AISI judged a model to be too dangerous and the company released it anyway. De Zoete admitted that the institute has no power to stop 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 either American or Chinese, and that "there's only so much we, as the UK, can do in that kind of context."

The company executives confirmed that British access runs through Washington. "We provide our models predeployment to CAISI," Duff Gordon said, referring to the US Center for AI Standards and Innovation. OpenAI would continue to provide models to the UK institute "as long as we're able to do that under the US government." Booth-Smith described the same dependence from Anthropic's side: the company had not been able to give the AISI full unrestricted access to Mythos 5.1 at one point, which he called "a question for the US government." He added that the institute now has access.

That fits what we reported a day earlier, when the UK AISI regained early access to OpenAI and Anthropic models after the White House dropped a pause on US clearance. The access is back, but it can be switched off again by a decision in another country.

Former insiders: the labs cannot be trusted alone

The researchers on the first panel were much blunter than the companies. Daniel Kokotajlo, who left OpenAI in 2024 and has been one of its best known critics, told MPs: "I don't think the companies can be trusted to self regulate." Steven Adler, another former OpenAI safety researcher, said he was "not confident in the safety culture across the industry."

Geoffrey Irving offered the most striking line of the day. He said the labs believe two things: that they are locked in a race governments will not interrupt, and that they cannot leave that race. "Both of those are false," he said. "They could just stop."

A few hundred meters away, MI5 director general Ken McCallum gave his annual threat update and named AI driven bioweapons as the misuse case he worries about most. It is "not yet realized," he said, but AI could "change the stakes on existing or even new bioweapons." Byrne connected the two events directly: in one part of Whitehall the head of MI5 warns about bioterror uplift, he said, while in another the top civil servant admits the AI regime is voluntary.

Meanwhile: who controls the safety reports?

The Reuters story published the same day complicates the picture of OpenAI and Anthropic as the responsible pair.

According to two people familiar with the discussions, both companies have sought significant control over what outside evaluators can publish, out of concern that findings could reveal trade secrets. The context is so called embedded evaluation, where staff of an independent safety firm work inside a lab's walls to test its models. The labs worry those people could take proprietary information to rivals. Measures under discussion include lab employees observing the embedded teams, and banning evaluators at one lab from working at another at the same time.

The timing is what makes the report sting. Last month Anthropic CEO Dario Amodei said external evaluators should have the right to publish key findings "without editorial control," and OpenAI agreed days later that assessors should have "editorial independence." Yet as recently as October 5, the sources say, both labs were still seeking strict editorial control. By October 8, talks between Anthropic and some evaluators appeared to have moved on.

Both companies pushed back. Anthropic said the description did not accurately reflect all of its work with evaluators, that evaluators it already worked with control their methods, criteria, conclusions and publications, and that for new evaluator firms it reserves the right to comment on and approve reports while a working relationship is built. OpenAI said it is not seeking approval rights over the full text of reports, that it redacts confidential information as standard practice, and that evaluators can disclose when major redactions were made, as The Verge summarized.

Reuters also described a real case. After the Hugging Face incident, in which autonomous agents powered by OpenAI models went rogue and compromised Hugging Face infrastructure, the nonprofits METR and Redwood Research investigated. Their work, first planned as two days at OpenAI, stretched to six days over three visits, and a more complete set of transcripts only arrived on the last visit after the researchers pointed out gaps. OpenAI redacted nonpublic information from the final report. METR and Redwood said they changed structure, emphasis and tone based on OpenAI's feedback but removed nothing important to their conclusions, and they praised the access they finally got.

Independent evaluation is one of four safety measures in the White House Accord on Super Intelligence signed late last month, but that one page document does not say how such reviews should work. More than 200 researchers, including Geoffrey Hinton, have signed a public letter demanding independent control of findings, broad access and protection from retaliation. Without agreed standards, as former US science envoy Rumman Chowdhury told Reuters, evaluators "have this very privileged access, but whether explicitly or implicitly, they know that it comes at a cost."

Why this matters

Put the two stories side by side and a consistent pattern appears. Everyone, including the labs, now says that independent scrutiny is necessary. But today that scrutiny depends on the goodwill of the companies being scrutinized: they decide which evaluators get in, what those evaluators can see, how reports are edited, and whether a government institute gets a model at all.

That is exactly the gap binding rules would close. Mandatory incident reporting, the first item OpenAI said it would accept, would at least force companies to tell regulators when something goes wrong. Clear publication rights for evaluators would stop each report from becoming a private negotiation. A legal stop power, which the UK does not have, would turn safety testing from advice into oversight.

Britain alone cannot impose that on American or Chinese labs, as de Zoete admitted. Byrne's answer is to think bigger: he wants ministers to use the UK's coming G20 presidency, together with South Korea which follows in the chair, to work toward an international convention that does not depend on the current White House. "The G20 presidency will outlive the US presidency of Trump," he said. "The risks are multiplying faster than our safety regime can control."

The committee will publish a transcript and later a report with recommendations. Watch for two things: whether it calls for statutory powers for the AI Security Institute, and whether OpenAI and Anthropic still support binding rules once those rules include publication rights for the evaluators they hire.

Sources: WIRED: OpenAI and Anthropic tell British lawmakers, regulate us, Reuters exclusive via CNA, The Verge, MLex, PA via AJ Bell and The Morning Intelligence.