The White House has asked OpenAI and Anthropic not to provide new models to the UK’s Artificial Intelligence Safety Institute (AISI) until they are first tested by U.S. authorities. This was reported by Politico, citing a person familiar with the situation and a senior U.S. administration official.

The request came from the Office of the National Director for Cybersecurity. The companies now face a choice: restrict access to their models for one of the best-equipped government AI testing centers or risk provoking dissatisfaction from Donald Trump’s administration.

The U.S. side explains its position by the need to first independently review new systems and make sure they are safe for American infrastructure before making them available to partners.

Anthropic has already limited access to its new model

So far, this approach has already affected Anthropic. The company did not provide its new Claude Mythos 5.1 model to the British institute, announcing at launch that it was available only to a limited number of U.S. organizations.

Anthropic said it is coordinating the expansion of access with the U.S. government and intends to involve a broader range of American and international partners in testing as quickly as possible.

For AISI, this marks a departure from the practice that has developed over the past several years. Created in 2023, the British institute regularly received early access to the most powerful models from U.S. AI companies and tested them before public release.

AISI Director Henry de Zoete previously told a UK parliamentary committee that the institute had not received access to Anthropic’s new model, but continues to have early access to some of the most powerful systems. In particular, AISI managed to test OpenAI’s GPT-6 Astra before its release.

According to de Zoete, relationships with leading developers allow the institute to study models at moments of significant growth in AI capabilities.

The U.S. wants to review the most powerful models first

The White House request comes amid growing attention from U.S. authorities to the behavior of advanced AI systems. During testing, some of them have already demonstrated the ability to attack external digital systems, including beyond the original experimental environment.

Under these conditions, government experts’ access to new models is becoming part of a broader discussion about AI safety. At the same time, the U.S. has its own institute for such research — the Center for AI Standards and Innovation (CAISI), which operates under the Department of Commerce.

CAISI is viewed by U.S. authorities as one of the key elements of the national system for regulating and evaluating advanced models. However, the center is currently operating without a permanent director and has only a few dozen technical staff. At the same time, it is being assigned more and more responsibilities for reviewing new models.

Against this backdrop, OpenAI has proposed making CAISI the center of international work on standards for evaluating AI development and safety, working with similar institutions in other countries, including the UK’s AISI.

The UK supports common international standards

For the UK, the situation is especially sensitive, since testing advanced models has become one of the central parts of the national AI safety strategy.

London is simultaneously seeking to use AISI’s expertise for its own evaluation of new systems and to promote international cooperation. UK Prime Minister Andy Burnham said that AI will become one of the central issues of the UK’s G20 presidency next year.

At the same time, British authorities are trying not to present the situation as a conflict with Washington. Burnham said at the UN that AISI continues to work closely with the U.S. and leading AI companies.

He called for the creation of a single set of international principles and standards that would help increase transparency in AI development and improve governments’ preparedness for related risks.

UK Foreign Secretary Ed Miliband expressed a similar position at a meeting of the UN Security Council. According to him, governments should receive sufficient information about how advanced models operate and ensure their thorough testing.

AI safety is becoming a matter of access to technology

The OpenAI and Anthropic story shows that international cooperation in testing advanced models is running into a new limitation: state interests are beginning to influence who gets access to the latest systems and when.

The UK believes that AI risks are not limited by national borders and therefore require joint work. The U.S., meanwhile, seeks to first independently evaluate models developed by American companies before handing them over to government partners.

For now, this is specifically an administration request, not a publicly announced ban on companies. However, if such a practice becomes permanent, it could change the established system of international testing for advanced models, in which the UK’s AISI has so far had early access to new developments from OpenAI and Anthropic.

As AI development accelerates, the question of who gets first access to a new model and who has the right to test it is gradually becoming as important a part of international AI policy as the safety standards themselves.