Reported 24 Sept 2026 Governance
The White House asked OpenAI and Anthropic to keep new models from Britain's testing agency until American testers are finished.
The Office of the National Cyber Director asked OpenAI and Anthropic to submit new frontier models to United States government testing before giving them to Britain's AI Security Institute, which has had pre-release access to American models since 2023. Anthropic has already done it: its newest model, Claude Mythos 5.1, went to a set of United States organisations only, and the institute's director confirms he has not received it.
The caveat that matters is that nobody has published the request, no date for it exists, and the only official explanation is one unnamed administration official saying this has been the policy for every new frontier model.
Frontier model safety testing before release was an international arrangement, with Britain's institute as the standing outside evaluator for American models.
Pre-release access is a national asset the American government allocates, and one lab has already stopped sharing on request.
Nothing here is a law or an order. A White House office asked two private American companies to sequence who tests their models first, and the reason given is that they are American companies. There is no statute behind it, no document, and no penalty for saying no. It would be easy to file this as a preference expressed in a phone call, and if Anthropic had shipped Mythos 5.1 to London anyway there would be no card. The reason there is a card is that Anthropic did not, and that the head of the British institute said so out loud.
What that establishes is a fact about power rather than about law. The 2023 arrangement worked because the labs wanted an outside evaluator with a government's reputation attached, and because no government objected. One government now objects, and the labs it objects to are the ones it licenses, buys from, regulates and, as the D.C. Circuit confirmed the next day, can designate. Britain gets whatever is left of the testing window after Washington is finished, and nobody will say how long that is. Pre-release access is a national asset the American government allocates, and one lab has already stopped sharing on request.
How it got here
- Jun 2023
OpenAI, Google DeepMind and Anthropic agree to give the United Kingdom early access to frontier models for safety research.
- Sep 15
OpenAI tells Congress it could accept outside inspectors written into federal law, and asks Britain to legislate on the same day.
- Sep 24
Politico reports the Office of the National Cyber Director asked OpenAI and Anthropic to hold new models from the British institute until United States testing is done.
- Sep 25
A British official confirms the request to Bloomberg. The institute's director says he has not received Claude Mythos 5.1.
- Who has to react
The AI Security Institute and Britain's Department for Science, Innovation and Technology, whose flagship capability now depends on an American queue; Henry de Zoete, who has to run an evaluator without the newest model, and Jade Leung, who leaves as chief technology officer at the end of the month; the European Commission's AI Office and every national regulator building an evaluation function on the assumption of pre-release access; the safety institute network agreed at Bletchley and Seoul, which has no answer to one member being cut off; OpenAI, which asked Britain to legislate nine days earlier and now has to explain the sequence; Anthropic, which complied while litigating against the same government; METR, Apollo Research and the other private evaluators, whose relative position improves if government access narrows; Chinese and other non-aligned labs, for whom this is an argument about whose testing counts; and Congress, where the OpenAI proposal for inspectors in federal law now reads differently.
The evidence
The request itself is not on the public record and no date for it has been given by anyone. Politico's Sophia Cai and Joseph Bambridge reported it on Thursday September 24, and a British official confirmed it to Bloomberg on Friday September 25. The White House did not respond to either outlet, OpenAI declined to comment to Bloomberg and Anthropic declined to comment to Politico, though Anthropic confirmed that Claude Mythos 5.1 is available only to a set of United States organisations. September 24 is the date the action became known, not the date it was taken, and we date the card accordingly. This surfaced late: September 24 sat inside Edition 032's window and we did not card it, so it is filed here at its true date rather than moved forward. Verification event: a named White House or Office of the National Cyber Director official confirming the policy on the record, or the United Kingdom's Department for Science, Innovation and Technology or the AI Security Institute stating in a published document or a parliamentary answer that pre-release access was withheld at United States request, by October 26, 2026.
ObservedReported Thursday September 24, 2026; confirmed by a British official Friday September 25. The White House Office of the National Cyber Director asked OpenAI and Anthropic to route new frontier models through United States government testing before sharing them with the United Kingdom's AI Security Institute.
The whole observation, as recorded
The named model is Anthropic's Claude Mythos 5.1, which the institute has not received; Anthropic said it is "only available to a set of U.S. organizations" and that the company is "coordinating with the US government to expand access to a broader set of domestic and international partners as quickly as possible." OpenAI's GPT-6 Astra had been tested by the institute before its release, which is the arrangement this changes. The only rationale on the record comes from a senior administration official: "Because they're American companies and this has been our policy with every new frontier model that comes out." Henry de Zoete, the institute's director, confirmed he had not received Anthropic's latest model and said the institute maintains "strong relationships with all frontier AI developers." Jade Leung, the institute's chief technology officer, is stepping back at the end of September. The arrangement being displaced dates to June 2023, when OpenAI, Google DeepMind and Anthropic agreed to give the United Kingdom early access to frontier models for safety research; that access is what made the institute, then called the AI Safety Institute, the most active pre-release evaluator outside the companies themselves. Neither government has published a document. The White House did not respond to Politico or to Bloomberg. Bloomberg's account on September 25 carried the headline that the President had told the two companies to withhold models from the United Kingdom agency. Nobody has said how long a United States review takes, which body conducts it, or what happens to a model that fails one. Separately and inside the same fortnight: OpenAI told Congress on September 15 that it could accept outside inspectors written into federal law and asked Britain to legislate on the same day, and Anthropic and OpenAI both committed on September 12 to giving outside evaluators desks and company laptops.
Established
- Politico reported on September 24, 2026 that the White House Office of the National Cyber Director asked OpenAI and Anthropic to withhold new frontier models from the United Kingdom's AI Security Institute pending United States government testing, and a British official confirmed the request to Bloomberg on September 25.
- Anthropic's Claude Mythos 5.1 was made available only to a set of United States organisations, and the institute's director, Henry de Zoete, confirmed he had not received it.
- A senior administration official said the request reflects standing policy for every new frontier model, on the ground that the companies are American.
- The institute had pre-release access to OpenAI's GPT-6 Astra, and the underlying access arrangement with OpenAI, Google DeepMind and Anthropic dates to June 2023.
Not established
- That any rule, order or contract term requires the labs to withhold anything. No instrument has been produced and the White House has not confirmed the request on the record.
- That OpenAI has complied. Only Anthropic's Mythos 5.1 is named, and OpenAI declined to comment.
- That the institute has been cut off. Its director says relationships with all frontier developers continue, and Anthropic says it is working to widen access.
- That American government testing is slower, weaker or stronger than the institute's. Nobody has published what the United States review consists of or how long it takes.
- Politico: White House asks OpenAI and Anthropic to hold new models from UK testers until US review, Sep 24
- Bloomberg: Trump tells OpenAI, Anthropic to withhold models from UK agency, Sep 25
- City AM: White House tells AI giants to hold models back from UK safety watchdog, Sep 25
- The Next Web: White House asks OpenAI and Anthropic to hold AI models from UK testers, Sep 25
Why it weighs this much
Score the act, not the alarm. Cost is low: an unwritten request from a policy office costs nothing and can be denied tomorrow, and the White House has not confirmed it to anyone on the record. Commitment is medium and only because of the official's sentence that this has been the policy for every new frontier model, which describes a pattern rather than a one-off, and because a company already acted on it. Reversibility is easy: there is no instrument to rescind, and Anthropic says it is working to widen access as quickly as possible, which is what compliance plus discomfort sounds like. Persistence is months rather than years, honestly, because the whole thing rests on goodwill in both directions and the next model release resets it. What keeps the strength off the floor is that a foreign government confirmed the effect on itself, which is the most expensive piece of evidence in the file and the one neither company volunteered.
What to watch
- strengthens A second frontier model is released without pre-release access for the United Kingdom institute, or the United States formalises a review requirement in an executive order, memorandum or contract term, by December 31, 2026.
- weakens The United Kingdom institute publishes an evaluation of Claude Mythos 5.1, or either government states that pre-release access has resumed on the previous terms, by December 31, 2026.
The words
- AI Security Institute
- The United Kingdom government body, formerly the AI Safety Institute, that tests frontier models for dangerous capabilities. Since 2023 it has had access to American models before public release.
- Office of the National Cyber Director
- A White House office created in 2021 to coordinate American cyber policy. It is a policy body, not a regulator, and has no statutory power to block a company from sharing a model.
- Pre-release access
- Giving an outside evaluator a model before the public gets it, so dangerous capabilities are found before they are deployed rather than after. Its value depends entirely on the evaluator getting the model early enough to matter.
- Frontier model
- The most capable models a company has, the ones whose new abilities are not yet understood. These are the only models pre-release testing is aimed at.
Your call
If the request is as reported, which explanation do you believe?
A call is final. Other people's calls show after yours.