White House tells OpenAI and Anthropic to withhold models from UK testers

By Splitfeed EditorialLast updated Sep 26, 2026, 19:17 UTC

Around September 24 to 25, 2026, the Trump administration directed OpenAI and Anthropic to withhold new AI models from UK safety testers until the United States government completes its own review. The directive, first reported by Politico, applies to both companies and effectively places American oversight authority ahead of any allied testing arrangement. The UK's AI safety agency, which had been in line to evaluate the models, cannot gain access until Washington clears the way. No specific models were named in the available reporting, though a separate article from around the same period referenced Anthropic's Claude Opus 5.5 launch, suggesting the directive may touch on recently released or upcoming systems.

The move carries weight because it disrupts a quiet assumption that had taken hold since the 2023 Bletchley Park summit: that the US and UK would coordinate on frontier AI safety evaluations rather than compete over them. By inserting a mandatory American review gate, the administration asserts a kind of jurisdictional primacy over models built by private US companies, even when those companies may have voluntarily agreed to share access with allied testers. Semafor characterised the White House position as isolated from broader AI safety cooperation efforts, framing it as a departure from the multilateral safety architecture that governments on both sides of the Atlantic have been trying to build. What is genuinely in dispute is whether this represents a reasonable assertion of national security prerogative or a move that weakens the overall credibility of AI safety infrastructure.

The pro-AI camp has not published reactions to this story yet. Typically, voices in that camp might argue that US-first review is a straightforward exercise of sovereignty over domestically developed technology, and that slowing deployment for additional government checks creates unnecessary friction. The anti-AI camp, represented here by the account Zubiqo, questions the practical credibility of the arrangement, noting that treating AI models like classified cyber weapons sits uneasily with the reality that the US agency now taking precedence lacks a permanent director and runs on only a few dozen staffers. The middle ground, reflected in a post from the account Grow My Bag, frames the story primarily through a global security lens, treating the access block as a significant geopolitical signal worth watching rather than an outright mistake.

The argument will sharpen once it becomes clear whether the US review process produces a formal mechanism with defined timelines, or remains an ad hoc gate that can be opened or closed at the administration's discretion. A permanent director appointment at the relevant US agency, or a public statement from OpenAI or Anthropic about how they are navigating the directive, would both clarify how durable this arrangement is likely to be.