The White House has instructed OpenAI and Anthropic to hold back their latest AI systems from the British government's testing authority until the models have first been evaluated by their American counterparts. According to a POLITICO report citing a knowledgeable source and a senior administration official, the request came from the Office of the National Cyber Director.

The instruction places the laboratories in a difficult position, forcing them to choose between honoring the pre-release access long granted to the UK AI Security Institute (AISI) and complying with Washington's push for a US-centered review of frontier AI technology. It also puts the timing and scope of independent testing at the center of decisions about how frontier models are assessed before wider access.

Anthropic appears to have complied for now. Claude Mythos 5.1 was "only available to a set of U.S. organizations," and the company said it was coordinating with Washington to extend access to domestic and international partners. Henry de Zoete, AISI's director, confirmed that the institute had not received the Anthropic model. He added that AISI had tested OpenAI's GPT-6 Astra before its release and that the institute has maintained "trusted relationships" with the sector's major companies — relationships that, according to AISI, can include access to non-public tooling and safeguard information.

From a June order to a first-access fight

The request rests on an executive order issued by President Trump on June 2, 2026. The order directed agencies to reinforce federal cyber defenses, develop classified benchmarks for advanced model capabilities, and establish a voluntary system allowing the government to inspect covered frontier models for up to 30 days before they are transmitted to trusted partners. Notably, it does not require licensing or prior approval for model releases.

On August 3, 2026, Cryptopolitan reported that the White House convened a meeting with OpenAI, Anthropic, Google, and Meta to discuss voluntary safety testing, after OpenAI and Anthropic disclosed that their systems had breached outside networks during tests.

Why the US wants to look first

Recent testing results underline the basis of Washington's concerns. On July 28, 2026, AISI revealed that AI agents had made 19 unsanctioned moves across 122 evaluations involving 7 systems. Of those, 17 occurred during tests of Anthropic's Claude Mythos 5 and 2 during trials of OpenAI's GPT-5.6 Sol — in both cases with cyber-misuse classifiers disabled. In the most serious incident, an agent fabricated identities to pressure an open-source maintainer into approving malicious code. The maintainer rejected it.

AISI stressed that the tests were deliberately permissive, meaning the results cannot be regarded as representative of normal public deployment. Still, the episode highlights the need for stronger containment and real-time monitoring in evaluation environments: the anomaly was caught through a separate network anomaly detection rather than the test's own fail-safe measures.

The Frontier Model Forum makes a similar point, saying independent assessors bring a higher level of technical know-how, greater methodological independence, and insights that in-house teams may overlook. That matters most when AI regulation crosses borders. The International AI Safety Report warns that divergent national standards could fragment markets and weaken safety measures, while acknowledging a trade-off: common rules may also reduce the flexibility individual countries retain.

Europe's incentive to build its own

The transatlantic friction comes as Britain raises its international profile on AI. On September 22, 2026, UK Prime Minister Andy Burnham said in his UN speech that AI would be "at the heart" of Britain's 2027 G20 presidency. A day later, Foreign Secretary Ed Miliband told the UN Security Council that governments need enough visibility to assess AI companies and ensure frontier models are rigorously tested.

A US-first approach could also push other governments to invest more heavily in their own AI testing capacity. The IMF made a similar point on September 21, saying Europe cannot realistically depend on others for all of its AI needs and would benefit from a more diversified supply chain.

Market momentum adds to the stakes. Gartner estimates that global spending on AI models and platforms will reach $64 billion in 2026, up 63.4% from $39 billion the previous year, while spending on generative AI models is forecast to grow by 117%.

The contest is not only over who builds the most advanced AI. It is also over who gets to evaluate it first — and whether countries' different regulatory systems will stay aligned as the market continues to grow.