White House Seeks US-First Review of OpenAI and Anthropic Models
Published by AINave Editorial • Reviewed by Ramit
The White House has reportedly asked OpenAI and Anthropic to let US authorities test new AI models before sharing them with the UK government’s AI Security Institute (AISI). The request, attributed to the Office of the National Cyber Director, puts a change in testing sequence at the center of a relationship where the UK institute had previously received early access to some models. Politico reported the request, citing a person familiar with the matter and a senior US administration official; the companies had not confirmed it in the cited coverage.
Early access is part of the testing arrangement
AISI has routinely tested some leading labs’ models before public release through voluntary arrangements. Its director, Henry de Zoete, said the institute had tested OpenAI’s GPT-6 Astra ahead of release and continued to have prerelease access to some of the world’s most capable models. That makes the issue more specific than whether a model is publicly available: the timing determines whether UK evaluators can examine it before launch. The institute’s prior access and GPT-6 Astra test were described by Politico.
Anthropic’s Claude Mythos 5.1 is a concrete example of access already being limited. The company did not provide that model to AISI, saying it was initially available to a set of US organizations, and said it was coordinating with the US government to expand access to domestic and international partners. That case should be kept distinct from the newly reported request: it does not establish that every model named in other coverage was withheld for the same reason, or that AISI has lost access to all new systems. Anthropic’s stated access position and AISI’s account appear in Politico’s reporting.
US review comes before sharing with partners
A senior US official said the administration wanted to review models and take steps to ensure US systems were secure before sharing them with partners. The reporting does not detail the review’s procedures or timetable, so the practical question is how long that first-look requirement would delay other evaluators. The stated rationale came from the senior official cited by Politico.
The sequence matters because AISI’s work supports the UK’s own response to AI safety risks and its stated ambition to help shape shared international standards. Meanwhile, the US Commerce Department’s Center for AI Standards and Innovation was facing a growing flow of frontier models with only a few dozen technical employees and no permanent director, according to Politico. OpenAI had argued that the US center should lead international standards work with other national institutes, including AISI. Those staffing constraints and OpenAI’s position were reported by Politico.
The reported request therefore sits uneasily beside calls for joint testing: US-first review could give Washington an earlier security check, while delaying another government’s independent evaluation. The UK government says it will continue working with the US and other partners. The reporting points to pressure on early access, not a confirmed end to cooperation. Whether the review becomes a short first step or a lasting condition will determine how much that pressure changes the shared testing model. The UK government’s stated commitment to continued cooperation was reported by Politico.
