White House cyber office asked OpenAI and Anthropic to hold new models back from UK safety testers

The claim originates with a single outlet. Politico reported that the Office of the National Cyber Director asked OpenAI and Anthropic to withhold their newest frontier models from the UK AI Security Institute's pre-release pipeline until US officials had completed their own review. A senior administration official was quoted giving the rationale in one sentence: because they are American companies, and this has been the policy with every new frontier model. Other outlets carried the story, but every version traces to that original reporting rather than to independent documentation, so the core allegation is recorded here as reported.

The specific assertion that could not be verified is the one about Claude Mythos 5.1 being restricted to a US-only group under Project Glasswing. Anthropic's own published Glasswing material does not support it. The programme page describes an April launch with a named partner list and more than forty further organisations; the June expansion post says the partner group spans more than fifteen countries and states an intention to expand geographically further. Neither page mentions Mythos 5.1 by version, and neither describes a US-only restriction. The version-specific claim is a reported one the company's own documents do not corroborate.

The on-record responses cut in interesting directions. UK AISI director Henry de Zoete did not deny a gap so much as change the frame, saying the institute maintains strong relationships with all frontier developers and continues to have pre-release access to some of the world's most capable models, naming OpenAI's GPT-6 Astra. Anthropic's cited language — that it is coordinating with the US government to expand access to a broader set of domestic and international partners as quickly as possible — reads as an acknowledgement that access is currently narrower than it will be, without conceding that an instruction caused it. Anthropic declined comment to Politico, OpenAI declined comment to Bloomberg, and the White House replied to neither.

The structural question is real whatever the precise facts. State pre-deployment evaluation of frontier models is voluntary and bilateral, built on lab discretion: there is no treaty, no statutory duty and no reciprocity requirement compelling an American lab to hand a model to a foreign safety institute. That architecture was always vulnerable to exactly this. If access becomes contingent on a US review step, the most capable evaluator outside the United States tests later, on older models, or not at all — a governance outcome achieved without anyone changing a rule.

• Reported Politico reported on 24 September that the Office of the National Cyber Director asked OpenAI and Anthropic to hold new frontier models from UK AISI pre-release testing until a US review completed. TheNextWeb
• Claimed That Claude Mythos 5.1 was limited to a US-only set under Project Glasswing. Anthropic's own Glasswing pages mention neither that version nor any US-only restriction, and say partners span more than fifteen countries. Anthropic
• Confirmed UK AISI director Henry de Zoete says the institute continues to have pre-release access to some of the world's most capable models, naming GPT-6 Astra. TheNextWeb