White House holds back models from London. AI testing becomes leverage

Access to frontier models has become an instrument of foreign policy. Politico reported that the White House asked OpenAI and Anthropic to stop passing new models to the British AI safety institute until the United States finishes its own assessment. The request came from the Office of the National Cyber Director.

A company between two governments

The British institute, the AI Safety Institute, is one of the best funded government testing bodies in the world, and until now it had early access to models. Putting it behind Washington in the queue leaves the companies in a difficult spot. Refuse, and they lose the relationship with London. Agree, and they displease the American administration. A senior US official says the policy is clear: the American assessment comes first, sharing with partners second.

Anthropic appears to go along with it. The company did not pass the Claude Mythos 5.1 model to the British institute, saying it can be shared "only with selected American institutions". It announced it would coordinate with the US government on widening access. Henry de Zoete, who heads the British institute, confirmed in a letter to a parliamentary committee that it has no access to Anthropic models. He noted, though, that the institute keeps earlier access to some models. OpenAI, for example, gave it GPT-6 Astra for testing before the official launch.

Who gets to write the standards

The dispute has a second, institutional layer. OpenAI argues that the American Center for AI Standards and Innovation (CAISI) should lead the creation of international standards for monitoring AI development and safety, working with national institutes, including the British one. CAISI, however, is an unfinished institution. It has no permanent head, and several dozen technical staff are supposed to assess an uninterrupted stream of new models. Congress and the White House are pushing to increase its resources.

Europe watches the dispute from a distance, but not without consequences. London is losing its position as a privileged reviewer. France did not sign the September declaration on strengthening oversight of frontier systems, while Germany and Spain did. At the same time OpenAI is preparing to present the GPT-6 Cyber model on 29 September at its developer conference in San Francisco. It is the company's fourth model this year intended for cybersecurity tasks, released under the Daybreak programme with two levels of access. Regulation, testing and standardisation are becoming one field of rivalry, in which models are at once a product and a bargaining chip.

The mechanism of pressure is subtle. Washington does not formally ban cooperation with London. It asks companies to hold back new models until its own assessment is done. The effect is the same. The testing institute stops being the first reviewer, and the British prime minister has to intervene by phone to get explanations. An ally's trust turns out to be a resource that can be managed just like access to chips.

Standards are becoming another export commodity. Whoever tests a model earliest shapes the technical requirements for the whole market. Whoever is shut out of that circuit gets the norms ready made. In this dispute the European Union has an asset in its own law, but its capacity to test frontier models is still smaller than the scale of the traffic it wants to assess.