White House tells OpenAI and Anthropic to hold new models back from Britain's AI Security Institute

For three years, the case for the UK's AI Security Institute rested on a simple claim: that Britain, allied with Washington but formally outside its chain of command, could examine the same frontier models the American government examined, at the same time, and catch what a single tester working alone might miss. On 24 September 2026, Politico's Sophia Cai and Joseph Bambridge reported that the White House has quietly ended that arrangement. A British official confirmed the report to Bloomberg the following day.

The request is narrower than 'the UK is banned from AI safety testing', and the precision matters. It does not touch models already released, and it is not a statute or an executive order — it is a policy position, relayed to two companies by an office few people outside Washington could name. The Office of the National Cyber Director sits inside the White House and ordinarily coordinates cyber policy; using it, rather than the Commerce Department's own testing body, to deliver the message tells its own story about how the administration is framing this — as a security instruction, not a science-sharing decision.

That is the entire stated rationale, attributed to a senior administration official. It is a jurisdictional argument, not a technical one: American companies' most capable systems get reviewed by the American government first, and only then do allies see them. Politico's sourcing — one senior official and one person familiar, both granted anonymity — means there is, as of this week, no public directive or memo to cite. The instruction exists because two companies are following it, not because a published rule compels them to.

The clearest evidence the request has teeth arrived before Politico's report did. Anthropic launched Claude Mythos 5.1 on 1 September 2026 and restricted it, as it had restricted the base Mythos model in June, to a set of US organisations. The company was explicit about the gap on its own product page, in language Politico later quoted directly in its report:

Read against Politico's reporting, that line stops being a routine access-tier disclosure and becomes the first documented instance of the White House's policy actually working. Anthropic did not contest it, did not negotiate a carve-out for AISI, and declined to comment when Politico and Bloomberg both asked. Given the company's history with this administration — a February order pulling federal agencies off its products, a June export-control letter that briefly locked all foreign nationals out of Mythos and Fable entirely — quietly going along is the least surprising part of this story.

The institute's public position is that its access is intact. AISI director Henry de Zoete made that case in a letter to the House of Commons Business and Trade Committee on 15 September 2026, responding to the committee's questions about the Mythos 5.1 gap:

He pointed to AISI's evaluation of OpenAI's GPT-6 Astra ahead of its public release as evidence the channel still works. Both things can be true at once: AISI still gets some models early, and it did not get one of the two most consequential releases of the autumn. A UK government spokesperson made the reassurance case in stronger terms to Politico, adding that 'these risks do not stop at national borders and no country can tackle them alone' — a fair description of the risk, but not, on its own, a rebuttal to the narrower claim that the US now sees new frontier models before the UK does.

What makes the sequencing sting is that AISI is not the junior partner on paper. Politico's reporting noted, almost in passing, that the US testing body Washington is now asking to go first — the Commerce Department's Center for AI Standards and Innovation (CAISI) — has no permanent director and only a few dozen technical staff, and is increasingly stretched as it is asked to evaluate an expanding stream of frontier models. AISI, by its own account, is considerably better resourced.

Since the Bletchley Park summit in November 2023, the working theory behind national AI safety institutes was that allied governments testing the same systems independently would catch more between them than either would alone — the same logic that underpins financial-market supervision and aviation safety. A US-first queue does not abolish that theory, but it does convert AISI from a parallel tester into a downstream one, seeing what Washington has already looked at, on Washington's schedule.

The politics of the same week make the gap harder to paper over. Prime Minister Andy Burnham used the UN General Assembly to describe AISI as working 'hand in glove with the U.S. and the leading labs on AI safety' and to call for a 'single set of global principles and standards'. President Trump, addressing the same UN gathering, dismissed any coordinated global framework as a 'globalist scheme'. Those are not compatible starting positions, and the model-access story is the first concrete evidence of which one is actually governing company behaviour.

Two dates now matter more than the headline. On 13 October, OpenAI, Anthropic, Google DeepMind and Meta are due before the Commons Business and Trade Committee — which had already been pressing the voluntary-access question before Washington's request became public — with de Zoete giving evidence the same day. And the next frontier release from either lab is the real test: if AISI receives it on a materially later schedule than CAISI, the 'nothing has changed' line will be difficult to repeat with a straight face.