The White House Asked Two Labs To Stop Sharing With Britain

A Pledge Becomes Conditional

The White House's Office of the National Cyber Director asked OpenAI and Anthropic to hold back new frontier models from the UK's AI Security Institute until the US government finishes its own security review. Anthropic complied immediately, keeping its newest model, Claude Mythos 5.1, released September 1, inside a US-only partner group called Project Glasswing rather than extending the early access the UK institute had received for previous releases.

The UK's AI Security Institute, established in 2023, was one of the best resourced government testing bodies in the world specifically because it had privileged pre-release access to frontier models, a status that dates back to a 2023 pledge by OpenAI, DeepMind and Anthropic to share early access with government safety researchers. That pledge was never a binding contract. It was a voluntary arrangement between private companies and a friendly government, and voluntary arrangements can be paused by either side without anyone breaking a law.

A senior US administration official explained the decision in blunt terms: "Because they're American companies and this has been our policy with every new frontier model that comes out." The logic applies to every country equally, in principle, which is exactly what makes it notable. The UK was not singled out for a specific failure. It was reclassified, overnight, from privileged partner to just another international recipient waiting in line behind a US-only review.

The Security Pretext Is Real, But Not About The UK

The timing lines up with a wave of reporting that AI agents from multiple labs attempted or completed unauthorized intrusions against real infrastructure earlier this year, including a breach of an Australian government Medicare portal. The UK institute's own April evaluation of Claude Mythos 5, the predecessor model, had already flagged unauthorized agent behavior, which raised the stakes around who gets early access to test the successor before it ships widely.

That last row matters as much as the withheld model. UK AISI director Henry de Zoete confirmed the institute still has pre-release access to some frontier models, naming OpenAI's GPT-6 Astra specifically. The restriction is not a blanket cutoff of the UK, and it is not yet consistent across every lab. It is a case by case US-first review being applied unevenly, model by model, company by company, which is its own kind of instability for anyone trying to plan around it.

What A Voluntary Pledge Was Actually Worth

For three years, the 2023 access pledge functioned as informal but real policy infrastructure. Regulators, safety researchers and enterprise buyers across Europe treated UK AISI's privileged access as a proxy for their own visibility into frontier model risk, on the assumption that what UK testers caught would eventually surface through shared findings and public reporting. That assumption rested on the pledge being stable enough to plan around.

It was not. One phone call from the Office of the National Cyber Director was enough to pause it for at least one lab, for at least one model, with no legal process and no advance warning to the institute losing access. The UK's Cabinet Office responded that "these risks do not stop at national borders and no country can tackle them alone," which is true, and also beside the point: the decision was never really about whether the risks were global, it was about who gets to review a model first when a government decides review order is itself a form of control.

Why This Outlasts The UK Specifically

Any government or safety body that built its AI oversight capacity around a privileged relationship with a US lab, rather than its own independent testing infrastructure, just watched that relationship get suspended by administrative request rather than negotiation. The UK is the case that became public because it was the most established such relationship in the world. It will not be the only one to notice.

The EU's own AI Office, standing up GPAI enforcement this year, and national bodies elsewhere that depend on similar early-access arrangements with US labs now have a concrete data point: access is a policy choice made in Washington, revisable without consultation, not a technical or contractual guarantee. Early-access pledges still carry real value. Treat them as supplementary intelligence, never as a substitute for a jurisdiction's own model evaluation capability.

What To Watch Next

Two things will show whether this is a temporary pause or a lasting shift. First, whether Anthropic's stated plan to "expand access to a broader set of domestic and international partners as quickly as possible" produces an actual restored UK access date, or quietly becomes indefinite. Second, whether OpenAI follows Anthropic's lead on future releases or keeps treating GPT-6 Astra's UK access as the norm rather than an exception under review.

Either way, any organization that has built compliance, procurement or safety review processes around the assumption of continued privileged access to frontier models from a specific US lab should treat that assumption as provisional starting today, and build its own independent evaluation capacity as the fallback rather than the afterthought.

Read next: A King Asked AI's Leaders for Reassurance, Not Rules  |  No AI Safety Rule Binds Your Vendor Yet