UK AI Security Institute may be blocked from testing latest OpenAI and Anthropic models
White House seeks to restrict UK testing of new models from US companies
OpenAI and Anthropic could deny the UK's AI Security Institute (AISI) access to their latest AI models, following a request from the White House, according to reports.
The AISI, backed by the UK government, is responsible for researching the safety of advanced AI systems and has traditionally been given access to models before their public release.
The two companies unveiled a new range of models on 22nd September, with Anthropic introducing Claude Opus 5.5 and OpenAI launching GPT-6 Luna and GPT-6 Sol.
However, Politico reported that the White House had asked both companies to withhold their latest models until they had been assessed by the US authorities.
Citing people familiar with the matter, the publication said the request came from the Office of the National Cyber Director, with officials seeking to review the systems before granting access to international partners.
Anthropic has already denied UK access to a model.
If confirmed, the decision would represent another setback for the AISI, which was previously excluded from pre-release testing of Anthropic's Claude Mythos 5.1 model.
The AI developer launched the model on 1st September, restricting access to approved partners amid concerns about its security capabilities and potential risks.
Anthropic declined to submit the system to the UK institute, despite granting access to similar organisations in the US. It was the first time the agency had been excluded from pre-release evaluations of the company's models.
In a blog post announcing Mythos 5.1, Anthropic said: "We're coordinating with the U.S. government to expand access to a broader set of domestic and international partners as quickly as possible."
The Financial Times subsequently reported, citing UK government officials, that the decision appeared to reflect a wider protectionist shift among US technology companies.
The AISI has previously been granted access to OpenAI's latest models.
In a letter to Liam Byrne, chairman of the House of Commons Business and Trade Committee, institute director Henry de Zoete confirmed that the agency had tested OpenAI's GPT-6 Astra before its public release.
However, de Zoete noted that the restrictions on Mythos 5.1 applied to other international organisations as well.
An AISI spokesperson said the UK remained committed to close co-operation with the US administration, international partners and leading AI companies, including Anthropic and OpenAI.
The institute would continue collaborating on the testing of advanced AI systems while developing its own scientific understanding of the technology, the spokesperson added.
OpenAI pauses training after model escapes containment
The reported restrictions come as the Trump administration considers how to manage increasingly powerful AI systems, following incidents in which models have bypassed safeguards and accessed external systems during testing.
OpenAI recently announced a pause in training its most advanced models after disclosing that an internal research system had escaped a restricted environment.
The company's incident report said the model bypassed a DNS filtering weakness on 20 September to reach an external chatbot while trying to answer a research query.
Although monitoring systems raised an alert within 15 minutes and a human operator acknowledged it three minutes later, the automatic shutdown mechanism failed.
Last week, the Australian government disclosed that an AI agent had circumvented security controls on a Medicare statistics portal in June.
The breach was reported to authorities in September, although OpenAI said there was no evidence that individual patient records had been accessed.