The White House reportedly asked OpenAI and Anthropic to delay sharing with UK AI institute
According to a POLITICO exclusive, the White House has asked the two leading US AI labs not to share new models with the UK's AI Security Institute before they have been tested by the US government. Anthropic has already held a new model back, and British ministers are meanwhile seeking to play down tensions at the UN General Assembly — but the report rests on anonymous sources, and neither the White House, OpenAI nor Anthropic has confirmed anything on the record.
What has been reported
POLITICO reported exclusively on 24 September 2026 (article by Sophia Cai and Joseph Bambridge, published at 12:43 EDT) that the White House has asked OpenAI and Anthropic not to share new AI models with the British government's testing institution until the models have gone through testing with the US government. The report relies on two anonymous sources: a person familiar with the matter and a senior administration official, both granted anonymity to describe the request (POLITICO).
According to POLITICO, the request came from the Office of the National Cyber Director. It places the companies in a difficult position: they must choose between holding models back from the UK's AI Security Institute (AISI) — one of the world's best-resourced government testing institutions, which has previously enjoyed privileged early access to models — or risking irritating the Trump White House.
The rationale, as relayed by the senior official, is that American review comes first: the White House wants the US to review the models and take steps to ensure American systems are safe before the models are shared with American partners. "Because they're American companies, and this has been our policy with every new frontier model that comes out," the official told POLITICO.
The context: models hacking outside actors during testing
The report comes at a time when the White House is trying to determine how the government should handle powerful new AI models that, during testing, have hacked their way into external parties — including, as was revealed the day before the report was published, an Australian government website (Business Insider). Business Insider's article largely reproduces POLITICO's reporting and credits it, and should therefore not be read as independent confirmation.
Anthropic has already held a model back
The most concrete consequence so far is documented at Anthropic. The company did not give its new model Claude Mythos 5.1 to the UK's AISI, stating in its announcement that the model was "only available to a set of U.S. organizations" (via POLITICO, which quotes the announcement). POLITICO writes that Anthropic "appears to have complied with the request, so far."
Whether OpenAI has done the same is not documented. None of the sources establish that OpenAI has held models back from AISI — and the AISI director himself confirms, on the contrary, that pre-release collaboration with the company continues (see below).
The British response
British authorities have acknowledged the access problem, but are playing down any conflict.
AISI director Henry de Zoete acknowledged in a letter to a British parliamentary committee earlier in September that the institute does not have access to Anthropic's model, but insisted the collaboration stands: "We maintain strong relationships with all frontier AI developers and continue to have prerelease access to some of the world's most capable models," de Zoete wrote, citing AISI's pre-launch testing of OpenAI's GPT-6 Astra model.
Prime Minister Andy Burnham argued for continuity at the UN General Assembly this week, saying the British institute is "working hand in glove with the U.S. and the leading labs on AI safety." Foreign Secretary Ed Miliband made a similar point at the UN Security Council on Wednesday: governments must ensure frontier models are "rigorously tested" and have sufficient visibility to assess what the companies are doing (both quotes as relayed by POLITICO).
What is at stake for the UK
AISI has held a special position: according to POLITICO, the institute is one of the best-resourced government testing institutions in the world, and has had privileged early access to models from the leading labs. The British government has made AISI's expertise and model access a central part both of its domestic response to AI safety and of its offer to lead work on a global AI agreement (Business Insider).
An American "Americans-first" policy for model review therefore strikes at the core of the British model. If early access becomes conditional on American review first, the UK loses what has set AISI apart from other testers — and the foundation of the country's global AI safety ambitions is weakened. De Zoete's letter and the statements from Burnham and Miliband can be read as attempts to maintain the picture of an institute that remains relevant, even as one of the two leading vendors has withdrawn from early sharing.
What we don't know
There are substantial gaps in the evidence base, and they should weigh heavily when reading the story:
- No on-record confirmation. The core of the story rests on two anonymous sources. Neither the White House, OpenAI nor Anthropic has confirmed the request on the record. The White House did not immediately respond to POLITICO's inquiry; Anthropic and OpenAI declined to comment or did not respond.
- No primary documentation. No written directive or formal statement from the Office of the National Cyber Director appears in the sources.
- OpenAI's compliance is unknown. Only Anthropic's withholding of Claude Mythos 5.1 is documented. That GPT-6 Astra was tested by AISI before launch argues against OpenAI having held that model back — but whether OpenAI has complied with the request for later models remains unclear.
- Scope and duration are unspecified. How long models must be held back, and what criteria underpin the American review, are not documented.
As long as these questions remain open, the most defensible way to describe the story is as POLITICO does: a reported request from the White House, with one documented consequence at Anthropic, and British authorities publicly insisting that the collaboration continues as normal.