White House restores UK AI Security Institute access to new models after a 17 day freeze

The White House has dropped its request that frontier labs hold new models back from the UK AI Security Institute, ending a freeze that lasted 17 days. The Telegraph reported the reversal on October 11, 2026, and outlets relaying it on October 12 credit lobbying by UK AI minister Kanishka Narayan. For anyone asking whether the UK AI Security Institute can test frontier models before release, the answer today is yes again, but the way it got there is the story.

What the freeze was

On September 24, 2026, Politico reported that the Office of the National Cyber Director had asked OpenAI and Anthropic to give Washington a first look at new frontier models before they went to the British institute, according to TNW's write-up of the report, published September 25. A British official confirmed the report to Bloomberg the next day.

The administration's explanation was short. A senior official told Politico: "Because they're American companies and this has been our policy with every new frontier model that comes out." City A.M. reported the same day that Washington wanted American systems secured before sharing them with allies.

The request was never published as a document. Everything known about it comes from Politico's reporting and the officials and companies who responded to it, so the exact wording and reach remain unconfirmed.

Where Anthropic landed

Anthropic's own announcement shows the effect. The page for Claude Mythos 5.1 states: "Currently, it is only available to a set of US organizations." It adds that the company is "coordinating with the US government to expand access to a broader set of domestic and international partners."

That is a primary-source statement of a restriction, and it is the clearest evidence in the public record that the freeze had teeth. Politico reported that Anthropic appeared to have complied and had not given Mythos 5.1 to the UK institute. Anthropic's separate October 6 post on its Cyber Verification Program says it reviews organizations in its specialized tier "in collaboration with the US government", which fits the same pattern of government-shaped access.

What the UK said at the time

AISI director Henry de Zoete wrote to a parliamentary committee in September, according to TNW, acknowledging that the institute lacked Anthropic's model. He said it had tested OpenAI's GPT-6 Astra before release and that AISI keeps "strong relationships with all frontier AI developers." Resultsense, summarizing LBC and The Guardian, adds that de Zoete said the institute could still assess models after release, and that testing before and after release is central to the UK's understanding of AI risk.

George Williamson, who now leads the Alan Turing Institute, told The Guardian that the UK "may not have access [to AI] in all the circumstances when we want it," calling that "not a place where a nation wants to be," per Resultsense.

What we do not know

The pages we opened leave several things open:

• When access actually resumes, and whether the models held back in the freeze, including Mythos 5.1, will now be handed over.
• Whether the change covers every lab or only OpenAI, Anthropic and Google DeepMind, which Startup Fortune lists as restored.
• Whether the US wants to test first. Startup Fortune says the administration's stated rationale was locking down American systems before sensitive capabilities went to foreign partners, which would still hold in principle.
• Why the White House changed course. The Telegraph credits ministerial lobbying; no US official has explained the reversal in anything we could open.

All of the reversal reporting traces back to one Telegraph report. We could not open the Telegraph itself, so treat the October 11 date and the lobbying account as reported, not confirmed.

Why it matters for the race

Pre-release testing is the one place governments see a frontier model before the public does. The UK institute has done this on a voluntary basis since 2023, and our explainer on the UK and US testing institutes sets out what each body actually tests. The freeze showed that the voluntary arrangement runs through company decisions, and company decisions can be steered by the government of the country the company sits in.

AISI's recent work shows what is at stake. Its cyber evaluations of frontier agents were paused and then restarted under tighter controls, which we covered in our report on the resumed evaluations. A body that runs those tests needs the model before launch to be useful; testing after release is better than nothing, but it cannot stop a launch.

There is also a counterweight. The US testing body, the Commerce Department's CAISI, has no permanent director and only a few dozen technical staff, per Politico as relayed by TNW. A rule that routes every model through Washington first would run into that capacity limit.

The testing relationship at a glance

Pre-release testing works by consent. A lab chooses to give a government institute a model before launch, usually under a voluntary arrangement, and the institute reports what it finds. The freeze did not change a law, because no law required the sharing. It changed what two companies were willing to do after a request from their home government. That is why the same arrangement can end in a single phone call and resume in another.

Our take

The reversal is good news and it is thin news. Access was cut by a request that no one published and restored by a decision no one has explained. Resultsense makes the point plainly: a request made once can be made again, so firms that lean on AISI evaluations should not treat restored access as settled.

What we would watch is simple. The next time OpenAI, Anthropic or Google DeepMind ships a frontier model, check whether AISI's evaluation lands before the launch or after it. If it lands before, the freeze was a one-off. If it lands after, the September request quietly became the rule. Either way, the UK's pre-release testing now has a known single point of failure, and it sits in Washington.