Frontier model safety evaluation depends on one prerequisite: developer cooperation with the testing body. When Anthropic declined to give the UK's AI Safety Institute, AISI, access to its latest model, it was the first time the British agency had been excluded from an Anthropic evaluation. Officials inside the British government now fear the decision signals a wider protectionist shift among tech groups.
AISI operates at a defined point in the pre-deployment stack. Its assessments require model access before a system reaches users. Without that access, the institute cannot perform its function. The British government built AISI on the assumption that frontier developers would grant that access; this case puts that assumption in question.
The concern inside Whitehall reaches beyond this single exclusion. The fear inside government is that major US-based developers are pulling back from voluntary access arrangements with foreign safety bodies. AISI's situation is described as an early instance of that pattern. The characterization inside government is of a wider protectionist shift among tech groups, placing this access refusal inside a larger question about whether AI developers will continue to cooperate with non-US safety infrastructure.
The institute's position as an independent checkpoint rests on being present at the evaluation stage. An access refusal removes it from the process for any model withheld. If the pattern holds, AISI's function becomes conditional on cooperation it cannot count on.
The exclusion from Anthropic's latest evaluation stands, on the record, as the first confirmed instance of that shift.