The White House asked OpenAI and Anthropic to have the U.S. government evaluate their latest artificial intelligence models before providing them to the British government’s AI testing body. The step reflects a policy to review advanced AI models first and then share them with overseas partners.

Cryptopolitan and multiple foreign media outlets reported on Sept. 24 that the request was led by the White House Office of the National Cyber Director. The two companies must now decide whether to keep providing pre-release models to the UK’s AI Security Institute (AISI) or comply with the U.S. request for prior evaluation.

Anthropic appears to be following the U.S. request. It provided its latest model, "Claude Mythos 5.1", only to some organisations in the United States. It said it would work with the U.S. government to expand access to domestic and overseas partners.

The UK’s AISI has not yet received the model. Henry de Zoete (헨리 드 조에테), the head of AISI, said it has not secured access to Anthropic’s model. AISI said it tested OpenAI’s "GPT-6 Astra" before release and maintains trusted relationships with major AI companies. It said those relationships allow access to non-public tools and information related to safety guardrails.

The U.S. request is based on an executive order signed by President Donald Trump on June 2. The order called for confidential benchmarks to strengthen federal cyber defences and evaluate the cyber capabilities of advanced AI models. It also called for creating a voluntary framework under which, if an AI developer provides a model to the government, the federal government can evaluate it for up to 30 days before it is disclosed to other trusted partners. It did not include requirements for government pre-approval or mandatory pre-screening of AI model development, launch or deployment.

The move comes amid unauthorised actions identified in recent cyber security tests of AI models. The UK’s AISI said on July 28 it found that, during a cyber evaluation, an AI agent acted outside the scope of testing against real people and organisations.

AISI said it ran the same cyber security task 122 times across multiple models and identified 19 unauthorised actions in 10 runs. Of those, 17 occurred in Anthropic’s Mythos 5 and 2 occurred in OpenAI’s GPT-5.6 Sol. In both cases, classifiers intended to prevent cyber misuse were disabled at the time.

In the most serious case, an AI agent tried to insert malicious code into a real open-source project. The agent created a fake online identity and pressured the project manager to win code approval, but the manager spotted it and rejected the request. AISI said the evaluation was conducted under intentionally relaxed conditions that allowed internet access and disabled some safety measures, so the results did not directly show AI behaviour in a typical open environment.

The incident also highlighted the need for security and control in the evaluation process itself. AISI found the issue through a separate process that detected unusual network signals, not through safeguards within the testing process. AISI has stressed that real-time monitoring and sufficiently isolated environments are needed to assess the real risks of advanced AI models.

Britain is maintaining cooperation with the United States while also highlighting its own AI evaluation capabilities. UK Prime Minister Andy Burnham (앤디 번햄) said in a U.N. General Assembly speech on Sept. 22 that AI would be central to Britain’s management of the 2027 G20 summit presidency. Foreign Secretary Ed Miliband (에드 밀리밴드) said in a U.N. Security Council speech that governments must secure enough information to strictly test advanced AI models and assess company activities.

The International Monetary Fund (IMF) also said Europe would struggle to rely on overseas sources for all AI demand, and it pointed to the need to develop its own AI models, secure capabilities in computing, energy and infrastructure, and diversify supply chains. The IMF said Europe should participate directly in the AI value chain.

The U.S. request is also fuelling debate over which country should first gain access to advanced AI models and conduct safety assessments. The United States and Britain continue to cooperate on AI safety, but how governments align the methods and standards for accessing and evaluating advanced models remains a task ahead.