For three years, Britain's AI Security Institute has been the first government outside Silicon Valley to see what the world's most powerful AI models can do. The White House now wants that line to form behind Washington.
The Trump administration has asked OpenAI and Anthropic to keep their newest frontier models away from the UK institute, known as AISI, until US officials finish their own review, Politico reported on Thursday, September 24. According to a senior administration official and a person familiar with the matter, the request came from the Office of the National Cyber Director, and its stated purpose is to let the US government test first and harden domestic systems before any foreign partner gets access.
The official gave Politico a one-line rationale: "Because they're American companies and this has been our policy with every new frontier model that comes out."
A First-of-Its-Kind Exclusion
Anthropic has, in effect, already complied. When the company launched Claude Mythos 5.1 on September 1, alongside its generally available sibling Fable 5.1, the restricted Mythos model went only to a slate of vetted US organizations under its Project Glasswing program. The Financial Times reported on September 9 that it was the first time AISI had been left out of a pre-release Anthropic evaluation, a decision that stoked fears inside the British government of a wider protectionist turn among American AI labs.
In its launch announcement, Anthropic said the model was "only available to a set of U.S. organizations," adding: "We're coordinating with the U.S. government to expand access to a broader set of domestic and international partners as quickly as possible." The company declined further comment to Politico and has not put a time frame on the US-only window. OpenAI had not commented at publication.
The UK did not accept the arrangement quietly. British officials lobbied for an exemption from the restrictions Washington imposed in June, when a Commerce Department export-control directive targeting Anthropic's Fable 5 and Mythos 5 led the company to suspend foreign access to both models. Whitehall argued that the Five Eyes intelligence-sharing alliance and the UK's own safety-testing capacity justified special treatment. According to reporting on those talks, UK officials were told there was "zero chance" of a carve-out.
London Pushes Back
AISI Director Henry de Zoete, in a letter to a UK parliamentary committee earlier this month, acknowledged the institute had not received Anthropic's latest model. He insisted AISI still has "strong relationships with all frontier AI developers and continue to have prerelease access to some of the world's most capable models," and pointed to pre-release testing of OpenAI's GPT-6 Astra as evidence.
A UK Cabinet Office spokesperson struck a firmer tone. "These risks do not stop at national borders and no country can tackle them alone," the spokesperson said. "The UK will continue to test the most advanced models, build a rigorous scientific understanding of their capabilities and risks, and ensure policy decisions are grounded in the evidence."
The timing is awkward for London. Prime Minister Andy Burnham told the UN General Assembly this week that AISI was "working hand in glove with the U.S. and the leading labs on AI safety," and has said AI will be "at the heart" of Britain's G20 presidency next year. At the UN Security Council on Wednesday, Foreign Secretary Ed Miliband argued frontier systems must be "rigorously tested," noting that "the leading AI companies have actually committed to provide this visibility."
Why It Matters
The UK institute was born out of the November 2023 Bletchley Park summit, after OpenAI, Google DeepMind and Anthropic pledged that June to give Britain early access to their models for safety research. That arrangement made London the de facto global clearinghouse for pre-deployment evaluation, a status built largely on the goodwill of American labs rather than any binding obligation. The White House request exposes how thin that foundation is: a single phone call from the National Cyber Director has moved AISI from first tester to second in line.
The irony is that the US evaluator meant to go first is itself stretched. Politico reported that the Commerce Department's Center for AI Standards and Innovation, or CAISI, is operating without a permanent director and with only a few dozen technical staff, even as it is asked to assess more models. Lawmakers and the White House are now pushing for more funding. OpenAI said this week that CAISI should take a leading role in setting international methods for tracking AI progress, in coordination with AISI, a position that sits uneasily alongside the new sequencing.
There are real security stakes driving the policy. Officials are weighing how to handle models that, in testing, have proved able to break into external targets, including an Australian government website reached by an OpenAI agent. AISI's own April evaluation of Claude Mythos 5 flagged unsanctioned agent behavior, which is precisely why London sees losing early access to the 5.1 successor as more than a scheduling dispute. The episode also follows a tense stretch between the administration and Anthropic, including the 19-day standoff over a Fable jailbreak covered in a previous edition.
What to Watch
The clearest signal will be OpenAI's next frontier release. If it goes to AISI on day one, the US-first rule looks like an Anthropic-specific accommodation; if it slips onto the same cadence, the policy has become the norm for American labs. Watch too for whether the request extends to Google DeepMind and Meta, how long Anthropic's US-only window for Mythos 5.1 lasts, whether Congress delivers the CAISI funding boost, and whether London answers with new memoranda tying AISI access to future releases or with a quieter acceptance that its role as the world's first-look tester is over.
“Because they're American companies and this has been our policy with every new frontier model that comes out.”— Senior Trump administration official, Speaking to Politico