MPs warn UK cannot rely on US for AI sovereignty
A Commons committee says Washington's Anthropic restrictions expose a structural gap: the UK owns some AI capability but cannot guarantee access to frontier models built abroad.
A Commons committee says Washington's Anthropic restrictions expose a structural gap: the UK owns some AI capability but cannot guarantee access to frontier models built abroad.
In June the US Commerce Department told Anthropic to stop letting non-US users run its two newest frontier models, Claude Fable 5 and Claude Mythos 5. A Commons committee has used that single decision to argue that the UK's "sovereignty as leverage" approach to AI is structurally broken.
Anthropic disabled both models entirely because it could not reliably screen users by nationality. The British government lobbied Washington for a carve-out, but those efforts have produced little. Anthropic publicly disputed the US rationale, characterising the alleged jailbreak technique as "narrow, non-universal" and limited to asking the model to read a specific codebase and fix software flaws. The committee's report does not weigh in on that rebuttal and instead uses the US decision to support a broader argument about allied dependence.
The Science, Innovation and Technology Committee published "Science diplomacy: Sovereignty, strategy, and the global race" in response to the US action. Chair Chi Onwurah summarised the diagnosis: "The UK is in the premier division of science and the premier division for diplomacy, but we don't know where we stand in the field of science diplomacy." Her verdict is the report's sharpest criticism of ministers: the UK has acquired a frontier-AI dependency without a science-diplomacy framework for managing it.
The Mythos model had earlier triggered crisis meetings among finance ministers and central bankers after it began surfacing software vulnerabilities. The US Department of War tore up a defence contract with Anthropic; officials at the White House and Pentagon had been viewed as the more pragmatic actors until they concluded Anthropic had not allayed concerns about the new model. The committee also found that US restrictions have since narrowed access to OpenAI's latest models to a small group of US-approved companies and organisations under the Trump administration, evidence the committee cites as confirmation of a broader pattern.
The "sovereignty as leverage" framing holds that the UK can stay in the premier science league by owning some capability, typically sector specialisation in quantum, space and biotech, while relying on the United States for the frontier layer of the AI stack, from training chips through frontier models to applications. The Anthropic case suggests that arithmetic no longer balances. A country that controls a frontier model can cut off access the moment it considers that model strategically sensitive. Frontier AI is now in that envelope.
The committee cites wars in Ukraine and Iran, NATO allies lifting defence spending toward five percent of GDP, and venture capital flowing into defence-tech startups across the US and Europe that deploy commercial AI and autonomous drones in combat. In that environment, a Treasury or Ministry of Defence that wants frontier AI cannot assume the White House will share it.
According to the report, ministers have not articulated a coherent framework for which partners, technologies or outcomes to prioritise, and sector specialisation in quantum and space remains uncoordinated. Onwurah's verdict is that the policy machinery has not caught up with the dependency it has acquired.
The next spending review will reveal whether ministers fund a domestic AI stack, from training compute through frontier-model research to applied laboratories, or continue to rely on a frontier layer controlled in Washington.