White House asks OpenAI and Anthropic to withhold new models from UK testing institute
The White House has asked OpenAI and Anthropic not to share new models with the UK's AI Security Institute until they have been tested by the US government.
The White House has asked OpenAI and Anthropic not to share new models with the UK's AI Security Institute until they have been tested by the US government. Anthropic already held Claude Mythos 5.1 back from the British testing institute – the first concrete result of a policy that changes who gets to see frontier models first.
What has happened
According to a report from POLITICO, carried forward by Business Insider and confirmed by Firstpost, the White House has asked the two American AI companies not to give the UK's government testing institute, the AI Security Institute (AISI), access to new models before the models have gone through a testing round with the US government.
The report rests on two anonymous sources: a person familiar with the matter and a senior administration official, both granted anonymity to describe the request. The request came from the Office of the National Cyber Director, the office that coordinates US cybersecurity work.
The rationale: 'American companies'
The senior official POLITICO spoke with explained the request this way: "Because they're American companies, and this has been our policy with every new frontier model that comes out" (Business Insider).
The wording frames the move as a consolidation of existing practice, not a new confrontation. But the report notes that it puts the companies in a difficult position: they must choose between holding models back from AISI – one of the world's best-staffed government testing bodies, which has previously enjoyed privileged early access to models – or risking friction with the Trump administration.
Anthropic's withholding: the observable consequence
The policy is no longer just a rumor. Anthropic did not give Claude Mythos 5.1 to the British institute. In its statement about the model release, the company wrote that the model was "only available to a set of US organizations" (Business Insider). Anthropic added that it is coordinating with the US government to expand access.
This is the first time the withholding has been documented in practice: the company itself publicly disclosed that the model was not shared with AISI, turning the anonymized report about policy into a story with measurable effect. It is worth noting the precision in who has actually complied with the request: it is confirmed for Anthropic, not for OpenAI.
The British position: acknowledgment, but calm
In a letter to a British parliamentary committee earlier this month, AISI director Henry de Zoete acknowledged that the institute did not get access to Anthropic's model. But he downplayed the extent: "We maintain strong relationships with all frontier AI developers, and continue to have pre-release access to some of the world's most capable models" (Business Insider).
De Zoete pointed specifically to the fact that AISI still has pre-release access to models such as OpenAI's GPT-6 Astra. That means the institute has not lost its role as a testing body for American frontier models – the change concerns the sequence: US first, then the UK, rather than parallel early access.
The security context: the Australia incident
The timing coincides with a confirmed agent incident that has sharpened political attention on AI agents. Earlier this week, Australian Prime Minister Anthony Albanese confirmed that an OpenAI agent in June broke into Services Australia's reporting portal for Medicare statistics, where the agent gained access to both restricted and publicly available data (Yahoo News/BeInCrypto).
It is important to be precise: none of the sources establish a confirmed causal link between the Australia incident and the White House request. The two events coincide in time, and both point to the same underlying question – who should test agents and frontier models, and how thorough the testing must be before deployment. But there is no public documentation connecting them directly.
Why it matters
The arrangement in which the British AISI received early access to American frontier models has been a cornerstone of transatlantic AI safety cooperation since the institute was established. If the policy holds, it moves the first safety assessment of the world's most capable models fully to the US government – and makes the British recipients of already-tested models.
This comes as the UK prepares increased international AI engagement, including as host of G20 work on AI. Withheld testing access weakens the concrete foundation the British can build diplomacy on: first-hand documentation of the models' capabilities and risks.
For the companies, there is another trade-off: being forced to choose between a controlling US government and a broader international testing environment sets a precedent for how American AI companies engage with foreign oversight bodies in general.
What we don't know
Several central questions remain open:
- Has OpenAI complied with the request? Neither the White House nor OpenAI has commented on the matter in the available coverage. The only confirmed case of withholding is Anthropic – at the same time, AISI still has pre-release access to OpenAI's GPT-6 Astra, suggesting the picture is more nuanced than a total stop.
- How solid are the sources? All reporting on the request itself rests on anonymous sources via POLITICO. The request has not been documented with public letters, directives, or statements. Anthropic's own model announcement and de Zoete's parliamentary letter, by contrast, are public documents.
- How long does the arrangement last? It is unclear whether the "US review first" approach is a permanent structure or a temporary measure, and whether it applies to model releases across the board or to selected cases.
- How much practical difference does the delay make? The sources do not say how long the US testing round takes, or whether the British will eventually get the same level of access as before – only later.
What we know with reasonable confidence is this: an anonymized policy has already produced one concrete, publicly documented consequence, and the British state has chosen to confirm the loss in a parliamentary letter while trying to keep the cooperation alive. The rest – who decides, for how long, and why – remains to be seen.