White House Asks Labs to Delay UK Model Testing
Washington reportedly wants OpenAI and Anthropic to give American reviewers first access, challenging the UK’s role in frontier-model safety testing.
The White House has asked OpenAI and Anthropic not to provide new models to the UK’s AI Security Institute until US government reviewers have tested them, according to reporting citing a senior administration official and a person familiar with the request.
A change to pre-release access
The UK institute has developed one of the strongest government testing relationships with frontier AI developers. It has evaluated models before public release and worked with US counterparts on capabilities including cyber operations, biological assistance, software development, and autonomous behavior.
The reported request would introduce a US-first condition for access to new systems from American companies. The White House position is that US authorities should assess the security of domestic systems before those models are shared with foreign partners. The request is not described as a public ban, and the reporting does not establish whether either company has formally accepted it.
The timing is significant. Pre-release access is scarce, and independent testing often depends on receiving models before broad deployment changes the risk profile. A delay can limit a tester’s ability to identify dangerous capabilities, compare systems across jurisdictions, or publish findings before users receive access.
An international governance fault line
The episode exposes a tension between national control and cooperative safety evaluation. Washington may view model access as a strategic asset; London has treated early testing as a way to provide independent scrutiny and strengthen international coordination.
The policy could also fragment the emerging testing ecosystem. If companies must satisfy one government before sharing models with another, evaluation becomes slower and more politically conditional. That matters beyond the US and UK because other regulators may face similar restrictions when seeking early access to globally deployed systems.
The request remains reported rather than codified policy, and the companies’ responses are not yet clear. But even an informal US-first rule would reshape who gets to inspect frontier models before the public does.