The UK is back to testing US models. Access was never a right

On Sunday, October 11, the White House reversed its September 24 request that US labs withhold pre-release AI models from Britain's AI Security Institute. Reuters reported the reversal via the Telegraph: American developers can again give the UK institute early access to models for safety testing without running afoul of the Trump administration. Spokespeople for the White House, OpenAI, and Anthropic did not immediately respond to Reuters' request for comment.

The freeze had lasted seventeen days. Politico first reported on September 24 that the Office of the National Cyber Director asked OpenAI and Anthropic to hold new models back from British testers pending a US national security review; a British official confirmed the request to Bloomberg the next day. The reviews were to run through CAISI, the US testing body.

The one concrete casualty is known. Anthropic visibly complied: while the pause was in force, UK officials were kept from reviewing Claude Mythos 5.1, which Anthropic restricted to US organizations. OpenAI's GPT-6 Astra had already reached British testers before the freeze landed, per the AISI's own letter to Parliament. So the memo interrupted some models and not others, depending on timing and on how each lab chose to read the request.

That letter, from AISI director Henry de Zoete to a parliamentary committee, is the sentence to remember. He wrote that the institute maintains strong relationships with all frontier developers and "continue to have prerelease access to some of the world's most capable models." Read it twice. Strong relationships are not access rights, and "some" is an honest way of saying not all. The freeze made visible what was always true: cross-border safety testing runs on voluntary arrangements that Washington switched off with a memo and switched back on after lobbying by UK AI minister Kanishka Narayan.

The restoration does not settle the direction of travel. It happened the same week the White House told AI companies that incident reporting is mandatory (the full story), and the same day Brad Smith named the UK institute as the body that should lead global AI rules. Today, Narayan is due to speak in London: reports say he will warn that the frontier labs must strengthen their security and cannot be left to regulate themselves, and announce amendments to the Cyber Security and Resilience Bill that would tighten requirements on critical-national-infrastructure companies to report AI-related security breaches. The government also plans to deploy frontier AI to protect Whitehall departments and to expand National Cyber Security Centre support for businesses facing AI-enabled threats. Those are reported plans ahead of the speech, not quotes from it.

Watch the asymmetry. The UK is moving toward binding reporting duties written into legislation, while the US mandate still rests on a statement and an MOU. And the most telling detail of the whole episode may be the smallest: one model, Mythos 5.1, demonstrated that a release can be geofenced at will. The labs can already change distribution in response to government pressure, in either direction. The freeze proved it in seventeen days. What nobody has answered yet is what happens the next time the memo is not reversed.