The United Kingdom’s ability to independently evaluate advanced artificial intelligence systems has come under strain following recent decisions by US companies and regulatory authorities, raising concerns about the implications for national security and technological sovereignty.

In June, export controls introduced by the US government allowed for the abrupt withdrawal of access to AI models without prior notice or contractual remedies. This move affected the British AI Security Institute (Aisi), which serves as a key institution for testing and monitoring AI technologies within the UK’s legal framework. More recently, Anthropic, a US-based AI company, declined to provide Aisi with pre-release access to its Claude Mythos 5.1 model. Observers indicate that this development marks a significant challenge beyond a missed opportunity for evaluation, highlighting the vulnerability of allied nations reliant on foreign AI providers.

The UK does not have the capital, infrastructure, or workforce to develop frontier-level AI models comparable to those in the US or China. Instead, its strength lies in managing the layer beneath the cutting edge—testing AI systems domestically and ensuring their safe deployment and operation under British law. Aisi embodies this capability but remains dependent on the cooperation of foreign technology companies and governments.

Critics argue that such dependency creates an asymmetry that could undermine national security. While businesses might endure temporary loss of access to specific AI models, governments require consistent ability to assess risks posed by AI systems increasingly integrated into critical infrastructure. Meanwhile, adversarial actors face no similar restrictions. The UK’s National Cyber Security Centre observed that within 18 months, public AI models advanced from limited functionality to successfully performing over half the steps needed to simulate attacks on corporate networks, at relatively low cost. Openweight models, which are freely available and can be modified to remove safeguards, further exacerbate this imbalance. Restricting access for allied evaluators does not, therefore, restrict potential attackers.

In response, experts suggest that UK ministers should engage with US policy makers to establish clear, published criteria governing AI model access, including advance notifications of restrictions, exemptions for allies, and priority access for those charged with national defense. This approach mirrors existing frameworks in areas like energy security and defense procurement.

Historical precedent supports this strategy. The uncertain reliability of the US GPS signal led Europe to develop the Galileo navigation system, while Russia and China revitalized Glonass and built BeiDou, respectively, investing tens of billions of dollars. Analysts warn that ambiguous or unexplained limitations on AI access could similarly incentivize allies to pursue costly independent capabilities.

The ongoing tensions underscore the delicate balance between global cooperation in AI development and the strategic necessity for allies to maintain sovereign oversight over technologies that underpin their security.