September 11, 2026, 15:44 GMT | Comment
The limitations of the UK’s reliance on voluntary arrangements for testing the safety of frontier AI tools were exposed this week, after reports that
Anthropic did not grant the AI Security Institute access to its latest model. Ministers said the AISI is continuing to collaborate with Anthropic, but experts have criticized what they call an overreliance on the institute for the UK's broader AI strategy. To tackle the limitations exposed this week, lawmakers have called for stronger powers and a statutory footing for the AISI, while diplomats stressed the need for closer collaboration with the EU on AI safety.
The limits of the UK's leadership on AI safety were tested this week by reports that Anthropic did not allow the UK's AI Security Institute to have access to its latest model — Claude Mythos 5.1 — for pre-release testing.
Even though these claims — first reported by the Financial Times — haven't been officially confirmed by either the company or the government, MLex understands that only a few US organizations currently have access to Mythos 5.1. This exposes the fragility of the voluntary frameworks on which the UK currently relies.
Some campaigners and AI engineers have called for stronger UK regulation on AI safety and for more powers to be granted to the AISI, so that it can formally request access to the latest models for testing. But other experts have argued that the UK is relying too much on the institute for its broader AI strategy.
Andrew Strait, an AI policy expert who recently left the AISI, where he was head of societal resilience, said the reports showed the UK had “hit the wall” on voluntary frameworks.
“If anything should become clear, it's that such frameworks fail, the minute dual-use security concerns and strategic advantage — or money and revenue — enter the picture,” Strait said. “The only way out of the broken logic of these market dynamics is government regulation that creates safety cultures, testing and release practices, and stronger liability requirements.”
Strait also warned that the US administration will seek to become the sole tester of these systems, and to push others out.
That concern was echoed in the UK Parliament this week. On Monday, senior lawmaker Lionel Tarassenko called for legislation to give the AISI statutory footing and broader powers, saying it is currently “just a government unit [that] could be abolished tomorrow" (see
here).
Keegan McBride, director of science and technology policy at the Tony Blair Institute, agrees that the voluntary nature of the AISI’s relationship with US companies is proving limited. But he said the UK should invest in its own AI capability, rather than make the institute's safety testing the center of its long-term AI strategy.
The AISI is a “fantastic institution,” McBride told MLex, but “not an institution that is going to provide you a strategic advantage over the US.”
If US companies do not share models for testing, “the AISI basically does nothing for you. If you’re not building or using the technology, you’re not relevant. ... Regulation is not enough. Being a regulatory superpower does not matter in this case. It's just not a path to meaningful competitiveness in the age of AI," McBride argued.
— Repeated warnings —
The reports that Anthropic refused the AISI access to its latest model came in the same week that warnings about the risks of superintelligent AI made headlines around the world, with former engineers at leading labs raising concerns about catastrophic risks to humanity.
Anthropic itself published a report on Thursday, documenting operations it disrupted between December 2025 and August 2026 involving threatening uses of its models. The company said misuse of AI is becoming more accessible and autonomous, with some operations running “autonomously, with minimal human input or supervision” (see
here).
But the timing of the superintelligent AI concerns raised this week has been questioned by industry voices: Anthropic is expected to issue its Initial Public Offering in mid-October.
Some experts say the latest warnings should be taken with a pinch of salt as they could increase Anthropic's prospects of being seen as ahead of the competition in its hopes to achieve superintelligent AI — potentially leading to higher share prices at its IPO.
On Monday, UK lawmakers, campaigners and policymakers were already discussing the harms of superintelligent AI at two separate debates in the UK Parliament that called for AI development to be subject to further controls (see
here). These calls included the presentation on Tuesday of a bill that seeks to ban the development of superintelligent AI in the UK (see
here).
These warnings then became louder, with one lawmaker taking the issue of Anthropic and the AISI to the
House of Commons' main chamber on Thursday.
Labour lawmaker Chi Onwurah, who chairs the Science, Innovation and Technology committee, said AI companies are debating about "the possibility of AI destroying humanity," and stressed "[while] the
European Union is moving ahead with a comprehensive AI regulatory framework, the UK continues to rely on existing regulation."
She asked minister Hamish Falconer what discussions the government was having with the EU about aligning AI regulation, "especially given that Anthropic has refused to submit its model to the AI Security Institute." Onwurah then asked Falconer how he would ensure the safety of European citizens.
The government avoided the EU question, with Falconer saying only that the government and AI minister Kanishka Narayan are "working at pace to address" the issues she raised.
— International collaboration —
Across the channel, calls for further UK-EU collaboration on AI were met with more enthusiasm. At an event* in Brussels on Thursday, UK diplomats said cooperation between governments, companies and the network of AI safety institutes would be needed as the risks from frontier models become more acute.
When questioned by MLex about the reports involving Anthropic and the AISI, Joshua Privett, head of digital at the UK Mission to the EU, said he couldn’t comment on this specific case. But he said the UK has a long history of cooperation with AI companies, while the EU has its own relationships, and that this dialogue needs to continue.
The Cabinet Office has made a similar point, saying the AISI continues to collaborate closely with industry partners, including Anthropic, and that it tested
OpenAI’s GPT-6 Astra before public release last week. The government also said that AI risks do not stop at national borders, and that “no country can tackle them alone.”
For Robert Trotter, economic affairs counsellor to the UK Mission to the EU, the recent “flurry of people resigning from frontier labs and talking about how dangerous the models are” demonstrates the importance of further EU-UK collaboration.
"That is an absolutely perfect example of how no single company can make frontier AI safe for everyone, and no single country can make AI safe for everyone. The only answer to AI security is multilateralism. Cooperation. There is no way around that," Trotter said.
He said this is one of the reasons the UK is "so committed and excited for an increased partnership agenda with the EU."
But the case for closer EU-UK cooperation does not answer McBride’s concern that the UK needs its own frontier capability. International partnerships can support testing and safety research, but they cannot replace access to the technology itself.
"If you're a country that is not building at the frontier, saying that you're regulating superintelligent AI is kind of irrelevant. You can do it, that's fine, but you have to be realistic about what that can and cannot achieve," McBride said.
* “Towards a UK-EU Partnership for a New Age” report launch, TechUK, Brussels, Sept. 10, 2026
Please e-mail editors@mlex.com to contact the editorial staff regarding this story, or to submit the names of lawyers and advisers.
Tags
Sections:
Artificial Intelligence
Industries:
Banking & Finance, Computing & Information Technology, Media & Telecommunications, Pharmaceuticals & Biotechnology, Computer Networks, Social Media, Digital Economy, Electronic Commerce
Geographies:
United Kingdom, Europe, Northern Europe