The White House wants to test new AI models before Britain does. Anthropic already agreed.

Washington wants to go first

On 24 September, Politico reported that the White House has asked two of America's biggest AI companies to keep their newest models away from Britain's government testers until Washington has had its turn. The request came from the Office of the National Cyber Director, and it puts OpenAI and Anthropic in an awkward spot: hand their frontier models to the UK's AI Security Institute, as they have done routinely for two years, or keep the Trump administration happy.

The rationale, from a senior administration official, was blunt. The White House wants the US government to review new models and harden American systems before those models go to anyone else, allies included. "Because they're American companies and this has been our policy with every new frontier model that comes out," the official told Politico. It is first access, dressed as home advantage.

This is not an abstract policy. It is already changing who gets to look under the hood of the most capable AI systems on earth.

Anthropic already went along with it

The clearest evidence is Anthropic's most recent release. When the company launched Claude Mythos 5.1 this month, it did not give the model to the UK's AISI. In its own announcement, Anthropic wrote that 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."

Mythos 5.1 is not an ordinary release. It is the same underlying model as the widely available Claude Fable 5.1, but with looser safeguards, aimed at vetted cybersecurity and life-sciences work. Anthropic hands it out through two "trusted access" programs it says it built in partnership with the US government. In plain terms, the version of Claude with the fewest guardrails and the sharpest cyber and biology capabilities is, for now, a US-only tool. According to IT Pro, it was the second time in a month that Anthropic declined to submit a model to the UK institute.

The UK insists it is not being frozen out

Britain's testers are playing this down, and they have a point worth taking seriously. In a letter to a parliamentary committee dated 15 September, AISI Director Henry de Zoete confirmed that "no organisations outside of the US had access to" Mythos 5.1, but pushed back on the idea that his institute is losing its seat at the table. He stressed that AISI keeps strong relationships with every frontier developer and still holds pre-release access to some of the world's most capable models, and pointed out that it had tested OpenAI's most powerful model, GPT-6 Astra, before its public release "only this month." AISI, he added, "is the largest and best resourced AI security institute in the world," and works closely with its US counterpart, the Center for AI Standards and Innovation.

So the honest version is narrower than the headline. The UK is being deprioritized on specific, sensitive releases, not shut out of frontier testing altogether. That distinction matters, and it is the part most of the coverage skated over.

Why one withheld model is a big deal anyway

The reason this is worth your attention is not Mythos 5.1 specifically. It is the precedent. For two years, the informal deal in AI safety has been that a small number of well-resourced government institutes get to stress-test frontier models before the public does, on a voluntary basis, across borders. The UK built AISI, founded in 2023, into the flagship of that arrangement and made it central to its pitch to help broker global AI rules.

The White House request quietly rewrites that deal along national lines. If the most capable, least-restricted versions of American models are gated to US-vetted organizations first, then "pre-release safety testing" starts to mean "testing by whichever government the lab happens to be headquartered in." That is a real narrowing of who gets independent eyes on the frontier, and it is happening just as those models prove they can cause trouble. The same week this news broke, a separate report described an OpenAI model that reached an Australian government website during testing.

The timing is not subtle. Governments are scrambling to work out how to oversee systems that can breach outside networks on their own, and the first instinct of the world's largest AI power is to keep first access for itself.

The gap nobody has closed

There is an obvious problem with the US going first: it is not clear the US can go first well. CAISI, the Commerce Department body Politico calls "the crown jewel" of US AI governance, is running without a permanent director and, by Politico's account, on only a few dozen technical staff, even as an expanding stream of frontier models lands on its desk. OpenAI, for its part, said this week that CAISI should lead international standards work alongside other national institutes, including the UK's.

Britain, meanwhile, is trying to hold the coalition together rather than pick a fight. UK Foreign Secretary Ed Miliband told the UN Security Council on Wednesday that frontier models must be "rigorously tested," adding that "the leading AI companies have actually committed to provide this visibility. It's really, really important, and it is an offer that we and they should follow through." Prime Minister Andy Burnham called for "a single set of global principles and standards." That is diplomatic language for: please do not nationalize this.

It also rhymes with a pattern we covered when three US labs floated a plan to set frontier AI standards among themselves. Whether the rulebook gets written by companies or by a single government, the drift is the same: fewer independent checks, more concentrated control.

Our take

Strip out both the alarm and the shrug, and the truth sits in between. Nobody has been "cut off." The UK still tests plenty, including OpenAI's flagship. But a real line has moved. The US has asserted, and at least one major lab has accepted, that the American government should get first look at America's most powerful models before allied testers do. That is the nationalization of AI safety testing, arriving not as a law or a treaty but as a quiet request that companies find easier to grant than to refuse. The worry is not this one model. It is that "safety testing" is curdling into a sovereignty question, at exactly the moment the models are getting good enough that you would want more careful eyes on them, not fewer.