The White House has asked OpenAI and Anthropic to delay giving the UK's AI Security Institute (AISI) pre-release access to two new frontier models until an American security review is complete, Politico reported Thursday and multiple outlets confirmed on Friday.
Two Models, One Pause
The affected releases were both scheduled for early evaluation with UK AISI staff – the arrangement the two labs have used since 2024 to hand British reviewers a pre-deployment look at frontier capabilities. According to the report, US officials have asked the labs to hold off until a domestic cybersecurity review can be run first, with the emerging Mathematical AI Safety Institute mentioned as one possible reviewer among US-based bodies.
Neither OpenAI, Anthropic, the White House nor UK AISI has commented on the specifics publicly. The White House framing, per Politico, is that American AI capability should get an American-led review before reaching foreign safety testers, even those from a treaty ally.
A Diplomatic Wobble For A Working Regime
The UK AISI has been one of the few functioning cross-border pre-deployment testing programmes, with UK reviewers having pored over models including GPT-5, Claude Opus, Gemini and others before release. Any US-imposed delay complicates a relationship that was expected to be broadened at next year's AI Summit, and comes on top of Washington's separate move to fold model access into export-control conversations.
What It Signals For Frontier AI Policy
The move sits alongside a wider slate of AI governance stories this week, including Anthropic's $11.6B Akamai compute agreement, OpenAI's rogue-agent report from Transluce and the emerging “SAFA” frontier-standards body being drafted by Google, OpenAI and Anthropic. Together they show a US administration keen to keep the biggest levers on frontier AI – model access, compute contracts and now international review – firmly in Washington's hands.
Reporting based on coverage from Politico and Quartz.