Anthropic Refuses to Provide New Models for Testing by UK AI Safety Institute
nashnova research
Anthropic released Claude Mythos 5.1 with pre-release testing restricted to vetted US institutions only, shutting out the UK's AI Safety Institute (Aisi) for the first time — a move that has alarmed British national-security officials and may signal a lasting shift in who gets to test frontier AI.
What happened?
Anthropic released Claude Mythos 5.1 without giving the UK AI Safety Institute (Aisi) pre-release testing access.
This is the first time the company has limited a new model's testing to vetted US institutions only.
By contrast, OpenAI still gave Aisi pre-release access before launching its latest model, Astra, this month.
Why is the UK government alarmed?
National-security officials worry this is not an isolated incident but a sign that Aisi could be systematically excluded from frontier-model testing going forward.
This means → the global standing Aisi built over nearly three years could erode fast if even one major lab changes policy.
The UK Cabinet Office last month ordered an urgent review of the threat posed by consumers and businesses losing access to the latest AI models.
What role do US export controls play?
In June, the US government imposed export controls on Anthropic's Fable and Mythos models after discovering jailbreak vulnerabilities — exploits that bypass AI safety guardrails.
The controls forced Anthropic to pull both frontier models from global markets for several weeks.
In plain terms = Washington first restricted Anthropic's global release rights; Anthropic then tightened overseas testing access — and the UK was caught in between.
Where does the Anthropic–US government relationship stand?
Washington had previously classified Anthropic as a supply-chain risk over disagreements on US Department of Defense contract terms.
The relationship has since improved, with the two sides collaborating closely on the Mythos and Fable launches.
Some UK government insiders believe Anthropic's refusal was at Washington's request, though this remains unconfirmed.
What comes next?
Anthropic said last week it is "coordinating with the US government to expand access to a broader set of domestic and international partners as soon as possible," but declined to comment on this specific case.
Whether Aisi regains pre-release testing access to Anthropic's frontier models will be a key marker for the direction of US–UK AI cooperation.
This reflects a broader shift: in the frontier-AI race, model-testing access is moving from a technical question to a piece of geopolitical leverage.
市场有风险，内容仅供研究参考，不构成投资建议。