What happened
In a U-turn reported 11 October 2026, the White House reversed its late-September directive (issued via the Office of the National Cyber Director) that had ordered OpenAI, Anthropic and other major labs to withhold their newest frontier models from the UK AI Security Institute (AISI) pending US review. After lobbying by UK AI Minister Kanishka Narayan, pre-release access to models from Anthropic (including Claude Mythos 5.1), Google DeepMind and OpenAI has been reinstated for British safety testing.
Why it matters
The original withholding directive — part of the US asserting national-security control over how frontier models are distributed for foreign safety testing — was a landmark cross-border AI governance action. Its reversal reshapes the practical balance of power in frontier-model oversight: the UK AISI regains the ability to safety-test the most advanced US models before release, and it reflects the high-level bilateral negotiation now governing international frontier-model access.
Action needed
Frontier-model developers should reconcile pre-release access commitments to the UK AISI and other international safety institutes with continuing US national-security directives; UK-based safety-testing and assurance teams should expect resumed US model access and plan evaluation pipelines accordingly.