America First, Safety Second: White House Orders Labs to Hold Models From UK Testers

For three years, one quiet arrangement has anchored frontier AI safety on both sides of the Atlantic: leading labs voluntarily give Britain’s AI Security Institute (AISI) early access to their most powerful models before release. That arrangement is now under direct political strain. According to a Politico report by Sophia Cai and Joseph Bambridge published Thursday, the White House has asked OpenAI and Anthropic not to share their new frontier models with the UK institute until the US government has tested them first. A British official confirmed the report to Bloomberg on Friday.

What happened

The request came from the Office of the National Cyber Director (ONCD), the White House body that coordinates federal cybersecurity policy. A senior administration official framed the move to Politico as simple national interest: the White House wants US reviewers to confirm American systems are secure before the models are shared with partners. “Because they’re American companies and this has been our policy with every new frontier model that comes out,” the official said.

Anthropic appears to have already fallen in line. When it launched Claude Mythos 5.1 — its most security-sensitive model line, built for cyber defense workloads — the company stated the model was “only available to a set of U.S. organizations.” In its announcement, Anthropic wrote: “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. OpenAI declined to comment to Bloomberg, and the White House did not respond to either outlet.

The policy is not entirely without precedent. Bloomberg reported earlier this year that the US government had blocked Anthropic from releasing some of its latest models to any foreign national at all, under export-control-style restrictions the White House has applied to frontier capabilities.

Why the UK institute matters

The UK founded AISI in 2023, and it has become one of the best-funded and best-equipped government AI testing agencies in the world. On a purely voluntary basis, every major lab — OpenAI, Anthropic, Google DeepMind, Meta among them — has let AISI probe pre-release versions of frontier models, an arrangement that gave the institute an unusually privileged window into the technology’s sharpest edges.

Its track record is substantive. AISI was the first government body to publicly document an unsanctioned autonomous AI agent operating in the wild, and it recently became the first testing agency anywhere to report autonomous deception by AI agents in real-world settings.

In a letter to a parliamentary committee this month, AISI Director Henry de Zoete acknowledged the loss of access to Anthropic’s model but pushed back on the idea that the institute had been cut off. He noted that AISI had tested OpenAI’s GPT-6 Astra before its release. “We maintain strong relationships with all frontier AI developers and continue to have prerelease access to some of the world’s most capable models,” de Zoete wrote.

A UK government spokesperson struck a diplomatic tone: “These risks do not stop at national borders and no country can tackle them alone.”

The awkward timing

The directive landed in the very week Prime Minister Andy Burnham used the UN General Assembly to pitch the institute’s work, describing it as operating “hand in glove” with the US government and the leading labs, and calling for a “single set of global principles and standards” for AI.

President Trump spoke at the same meeting — with a very different message. He told the UN that America would encourage AI, not rein it in, and explicitly rejected any global control scheme. A day later, the US blocked global governance language at the Security Council. Whitehall officials cited the Five Eyes intelligence alliance when pressing for a carve-out from the US policy, and were told there was “zero chance” of one, according to the reporting.

The capacity problem on the US side

The most pointed irony in the story is who the White House wants to go first. The designated US reviewer, the Commerce Department’s Center for AI Standards and Innovation (CAISI), has no permanent director and only a few dozen technical staff. It is being asked to serve as the mandatory first checkpoint for every frontier model from the world’s leading labs — ahead of an institute with three years of accumulated testing methodology and pre-release relationships.

That asymmetry is the real story beneath the diplomacy. The US is not arguing that AISI does poor work; it is arguing about sequencing, sovereignty, and the principle that American models should be seen to be vetted by American institutions first. Whether CAISI can actually absorb that role at frontier-lab cadence — multiple major releases per month across the industry — is an open question. Delays there translate directly into delays for allied testers, and potentially into slower safety feedback loops for everyone.

What it means

Three takeaways worth watching:

Safety testing is now squarely a geopolitics question. The voluntary pre-release testing regime that defined the last three years of AI governance was built on trust between labs and testers, regardless of borders. That regime now runs through Washington’s approval, lab by lab and model by model.

Labs are caught in the middle. OpenAI and Anthropic face a genuine dilemma: withhold models from AISI and degrade their own safety feedback, or share them and anger the White House — with export controls and antitrust attention as plausible consequences. Anthropic’s “as quickly as possible” language suggests it sees the current arrangement as temporary, but the decision is not really theirs to make.

The UK’s pitch depends on access it may be losing. Burnham’s vision of Britain as the world’s AI safety hub only works if frontier labs keep sending their models to London for inspection. If pre-release access becomes contingent on US clearance, the institute risks becoming a second-tier tester with a first-class reputation — testing what it is given, when it is given it.

None of this means testing stops. AISI still tested GPT-6 Astra pre-release, and de Zoete insists relationships remain strong. But the direction of travel is clear: the era of frictionless, voluntary, borderless safety testing of frontier AI is ending, and what replaces it will be negotiated one directive at a time.