UK Wins Back Access to Test US AI Models

The UK government has confirmed that its AI Security Institute (AISI) is evaluating the latest generation of advanced artificial intelligence models, following talks with the United States over restrictions that had disrupted British access to technology developed by leading American companies.

AI Minister Kanishka Narayan confirmed on Monday that researchers at the institute were conducting safety assessments of new systems, including Google's Gemini 4 Argon. The announcement follows reports in September that the White House had asked OpenAI and Anthropic to withhold unreleased models from British researchers pending a US government review.¹

The restrictions raised questions about Britain's ability to independently assess increasingly powerful AI systems, particularly those capable of identifying software vulnerabilities, conducting cyber operations or carrying out complex tasks with limited human supervision.

Speaking on Monday, Narayan said the institute was now undertaking evaluations designed to identify risks before advanced models become widely available. The minister attributed the progress to discussions with American officials and continued cooperation with AI developers.^1,2

The development marks a significant step for the UK government, which has sought to establish Britain as an international centre for AI safety research despite not being home to the companies developing many of the world's most advanced models.

The institute, originally established as the AI Safety Institute following the UK's 2023 AI Safety Summit, works with leading AI developers to examine the capabilities and potential risks of frontier models. Its evaluations include assessing whether systems could facilitate harmful activities, expose cybersecurity vulnerabilities or behave in unexpected ways.

Under voluntary arrangements with developers, the institute has previously received privileged access to models before their public release. This allows researchers to identify potential weaknesses and provide findings to developers while changes can still be made.

However, the arrangements also mean that access depends on the cooperation of technology companies and, in some circumstances, decisions taken by the governments in which those companies operate.

The difficulties became particularly apparent earlier this year when the US government introduced restrictions affecting Anthropic's most advanced models.

On 12 June, Anthropic confirmed that it had received a US government directive restricting access to its Fable 5 and Mythos 5 systems by foreign nationals, including those located within the United States. The company said the restrictions had forced it to suspend access to both models for customers while it sought to comply with the directive.

The US government cited national security authorities. According to Anthropic, the directive did not provide detailed information about the specific concern, although the company understood it to relate to a potential method of bypassing safeguards designed to prevent misuse.

Anthropic disputed the justification for the restrictions, arguing that the technical capability demonstrated in the reported incident was already available through other AI systems. The company nevertheless complied with the government's instruction.³ The restrictions were lifted on 30 June, and Anthropic restored access to both models on 1 July.

The decision prompted questions in the British Parliament about the implications for national security and the country's ability to conduct independent AI evaluations.

In June, members of the House of Lords raised concerns about restrictions on foreign access to Anthropic's technology and asked what alternative domestic capabilities or international partnerships the government was pursuing.⁵

The issue returned to Parliament in September, when ministers were questioned about Anthropic's decision not to provide the AI Security Institute with pre-release access to its latest frontier model.

Responding on 18 September, Narayan emphasised that the institute maintained voluntary relationships with major AI developers and was in regular contact with Anthropic and other industry partners. He said previous privileged access had enabled researchers to identify dozens of vulnerabilities that developers subsequently addressed before releasing their models.⁴

Further reporting in September suggested that the White House had asked American AI companies to delay sharing new models with British testers while US officials reviewed access arrangements.⁷

Although the UK government did not publicly confirm that the institute had lost access to every leading American model, the reports highlighted the potential vulnerability of a system of international safety evaluation that relies on agreements with companies based overseas.

Monday's announcement indicates that the immediate dispute has been addressed sufficiently for the institute to resume evaluating the latest generation of models.

Narayan expressed confidence that the UK would continue to receive access following discussions with American officials, while acknowledging the uncertainty inherent in depending on technology controlled by organisations outside Britain.

The development comes as the government considers strengthening its approach to AI safety and cybersecurity. In remarks reported by the Financial Times on Monday, Narayan outlined proposals for compulsory safety reporting and monitoring of advanced AI systems, alongside the possibility of requiring developers to provide models for evaluation before release.⁶

Such measures remain under consideration rather than established legal requirements. Ministers are also weighing the potential implications for international relationships and cooperation with American technology companies.

The government is particularly concerned about the cybersecurity capabilities of increasingly autonomous AI systems, which can carry out sequences of tasks and interact with external software with less direct human involvement than conventional chatbots.

Recent disclosures by AI developers concerning unexpected behaviour during testing have intensified scrutiny of how these systems are evaluated and controlled. Narayan warned that advances in AI could make some forms of cyber intrusion more effective and called on developers to investigate serious risks before deploying new models.

The National Cyber Security Centre and the AI Security Institute are also expected to develop approaches for responding to incidents involving autonomous AI systems, including ways to contain harmful activity and support affected organisations.²

For the cybersecurity industry, the dispute over model access illustrates a practical difficulty in assessing emerging technologies. Independent testing depends not only on the expertise of researchers but also on whether they can examine systems before deployment, when potential weaknesses can still be addressed.

The UK has invested in developing that expertise and has established relationships with several of the world's leading AI laboratories. However, it does not control the development or distribution of most frontier AI models, leaving aspects of its evaluation programme dependent on international cooperation.

The latest agreement appears to have restored an important part of that cooperation, although the government has not publicly identified every model currently available to British researchers or established that unrestricted access will continue indefinitely.

The distinction matters as governments consider whether voluntary agreements provide sufficient oversight of increasingly capable AI systems or whether more formal requirements will be necessary.

For now, the AI Security Institute has confirmed access to at least part of the latest generation of frontier technology, allowing its researchers to continue work on identifying security risks before those systems reach the wider market.

The episode has nevertheless demonstrated that the ability to evaluate advanced AI is not determined by technical expertise alone. It also depends on the commercial and government decisions governing access to the models themselves.