Politico: the White House asked OpenAI and Anthropic to keep new models from Britain's AI Security Institute until the US has reviewed them

The White House has asked OpenAI and Anthropic not to give their new models to the UK's AI Security Institute until the US government has tested them first, Politico reported on 24 September 2026. The request came from the Office of the National Cyber Director. Anthropic appears to have gone along with it. OpenAI had not commented.

It lands on AISI, the British government's testing body, and on an arrangement the UK government has made, in Politico's words, "a central part of its domestic response to AI safety concerns". It started as a promise. On 12 June 2023 Rishi Sunak told London Tech Week that Google DeepMind, OpenAI and Anthropic had "committed to give early or priority access to models for research and safety purposes", as TechCrunch reported it at the time. That promise became the institute's routine. Politico describes AISI as "routinely testing their latest models ahead of public release on a voluntary basis". Voluntary cuts both ways. A lab that volunteered can be asked by its own government to wait.

The policy is sourced to people who would not be named

Politico's reporters, Sophia Cai and Joseph Bambridge, got the request itself from "a person familiar with the matter and a senior U.S. administration official, both of whom were granted anonymity". The official gave the reason: the White House wants the US to "review the models and take steps to make sure U.S. systems are secure" before they are shared with partners. And the scope: "Because they're American companies and this has been our policy with every new frontier model that comes out."

That last sentence is the strongest claim in the story and it has no name on it. The White House did not immediately respond to Politico. Neither did OpenAI. Anthropic declined to comment. Politico does not say when the request was made.

What is on the record says the gap is real

Anthropic's own page for Claude Fable 5.1 and Claude Mythos 5.1, which an earlier post here read in launch week, says Mythos 5.1 "is only available to a set of US organizations, though we're coordinating with the US government to expand access to a broader set of domestic and international partners as quickly as possible." Politico reports that Anthropic did not provide Mythos 5.1 to AISI. Nothing fetched says the White House request is why the page reads that way, and it may not be.

AISI's director, Henry de Zoete, wrote to a UK parliamentary committee earlier in September. In Politico's account he acknowledged the lack of access to Anthropic's model and said: "We maintain strong relationships with all frontier AI developers and continue to have prerelease access to some of the world's most capable models." His example was OpenAI's GPT-6 Astra, which AISI tested before its release. "Some" is the word to hold on to. Sunak's 2023 announcement had no "some" in it.

The Astra example also sits oddly beside the official's "every new frontier model". Either Astra went through a US review before AISI saw it, or the policy was not applied to it. Neither Politico nor anything else fetched says which.

The British government's line, from a spokesperson quoted by Politico, is that AISI "continues to work closely with the U.S. government, international partners and leading AI companies, including OpenAI and Anthropic." At the UN General Assembly the same week, Prime Minister Andy Burnham called the institute's relationship with the US and the labs "hand in glove".

The US queue is a small agency

Going first means the American reviewer sets the pace. The site's OpenAI entry records the 2 June 2026 executive order that set up voluntary pre-release access of up to 30 days for federal agencies. Politico does not name the agency that would do the testing the request asks for. It does describe the one that evaluates frontier models, the Commerce Department's Center for AI Standards and Innovation: CAISI is "operating without a permanent director" and "with only a few dozen technical employees" while it is asked to evaluate "an expanding stream of frontier models". If every new frontier model from the two labs waits on that team before London sees it, and release dates do not move, AISI's pre-release window shrinks by however long the US review takes.

Pre-deployment testing by more than one government, as in the joint US–UK look at o1 in 2024, only ever ran on access the developer chose to grant. What Politico adds is that the developer's own government now has a say in who gets it, and in what order.

OpenAI's answer is the one to watch. Politico notes that OpenAI said this week that CAISI should lead international standards work, alongside other national institutes including AISI. That is not an answer to the request, and in the coverage available on 25 September it had given none.

Sources

Retrieved 25 September 2026.

• Sophia Cai and Joseph Bambridge, White House asks OpenAI and Anthropic to hold new models from UK testers until US review, Politico, 24 September 2026. The declared anchor. Politico's own page returned HTTP 403 to direct retrieval, so every quotation above was read from Politico's copy on Yahoo News, timestamped 4:43 PM UTC on 24 September, whose canonical link is the Politico URL.
• Anthropic, Claude Fable 5.1 and Claude Mythos 5.1: the "set of US organizations" sentence.
• Natasha Lomas, OpenAI, DeepMind and Anthropic to give UK early access to foundational models for AI safety research, TechCrunch, 12 June 2023: Sunak's announcement.

De Zoete's letter is published on the UK Parliament's committees site as a PDF. On 25 September that site answered this author's fetches with a Cloudflare challenge page instead of the document, a block on this author's access path and not on the letter, which is public. The letter was not read here, so his words above are as Politico quotes them.