White House Tells OpenAI and Anthropic to Hold New Models From UK Testers Until US Review

The United States wants to see America's most powerful AI models before its closest ally does. The White House has asked OpenAI and Anthropic not to share new frontier models with Britain's AI Security Institute until a US government review is finished, according to Business Insider.

For years, the UK institute has been one of the first outside bodies to test new models before release. That arrangement now comes with a queue, and Washington is at the front of it.

Who is asking, and why

The request came from the Office of the National Cyber Director, the White House office responsible for national cybersecurity strategy, rather than from the Commerce Department, which runs the government's own model testing center. That detail matters. It frames frontier AI as a question of defending US networks first, and of sharing with partners second.

A senior administration official put it plainly: the US reviews the models first and secures American systems before anyone else gets them. The official described it as the policy for "every new frontier model" built by American companies.

It is a request, not a law. There is no statutory export ban and no formal break with the UK, and nothing described so far bars allies permanently. But when the White House asks two labs that depend on federal goodwill, the ask carries weight.

Anthropic appears to be complying

Anthropic already seems to be following the new order of operations. It announced Claude Mythos 5.1 as "only available to a set of US organizations," and said on its blog that it is coordinating with the US government to expand access later.

The UK side has confirmed the gap. Henry de Zoete, director of the AI Security Institute, told a parliamentary committee the institute did not have access to Anthropic's new model. He added that it still gets prerelease access to some systems, including testing OpenAI's GPT-6 Astra ahead of its launch.

London plays it down

British leaders are not picking a fight. Prime Minister Andy Burnham and Foreign Secretary Ed Miliband both downplayed any clash, and a government spokesperson said the institute still works closely with the US and with the labs, Reuters reported.

The awkward part sits on the American side. The Center for AI Standards and Innovation, the Commerce unit expected to carry out much of this review, has no permanent director and is widely described as short on staff. Washington is claiming first look at models it may not have the bandwidth to examine quickly.

The hacking backdrop

The timing follows reports that powerful models have attacked outside systems during testing, including an Australian government website cited by Business Insider. Officials say that is why they want eyes on the models before anyone else.

It also fits a wider pattern of governments treating frontier models as strategic assets. China's internet regulator is running its own probe involving DeepSeek, Moonshot and Claude. The difference here is that the country being made to wait is one of America's oldest allies.