For 17 days, Britain's main AI testing body was locked out of the newest American frontier models. On Oct. 11 that order flipped. The White House reversed a Sept. 24 directive that had told major U.S. labs to stop sharing unreleased systems with the UK's AI Security Institute, restoring pre-release access for safety evaluation, The Telegraph reported and Reuters confirmed.
The freeze came from Donald Trump's Office of the National Cyber Director. It asked companies including OpenAI and Anthropic to hold new models back from British testers until Washington finished a national security review of the arrangement. Anthropic complied in a visible way: UK officials were kept from reviewing Claude Mythos 5.1 while the pause was in force. The restoration, according to the Telegraph's reporting, followed lobbying by AI minister Kanishka Narayan and other ministers. AISI is again expected to receive early access from Anthropic, Google DeepMind, and OpenAI.
That is a short timeline for a national-security justification. Seventeen days is long enough to prove the United States can interrupt allied testing by memo, and short enough to make the interruption look reversible under political pressure. Spokespeople for the White House, OpenAI, and Anthropic did not immediately respond to Reuters when asked for comment.
AISI is not a symbolic office. Founded in November 2023 ahead of the Bletchley Park AI Safety Summit as the AI Safety Institute, it was renamed the AI Security Institute in February 2025 to sharpen its focus on cyberattacks, fraud, and bioweapon risk. The United States and the United Kingdom had signed a formal agreement to jointly test frontier models and share findings. The September order briefly suspended the practical side of that partnership. The October reversal puts the paperwork back in motion.
The episode lands against a louder British push for rules that do not leave frontier labs to police themselves. Narayan is set to announce measures tied to Labour's Cyber Security and Resilience Bill, including stronger duties for firms on critical national infrastructure to report AI-linked security breaches, plus more National Cyber Security Centre support for companies facing AI-powered attacks. At the United Nations, Andy Burnham has framed Britain as an "honest broker" for global principles on frontier AI and pointed to the UK's coming G20 presidency. Those ambitions looked thinner while London could not see the models Washington was still shipping to American testers.
Trump's public line on AI has been expansionist. At the UN General Assembly he said he wanted to "encourage" AI, "not rein it in," and warned against "globalist" interference with American companies. The freeze, and then the unfreeze, sit awkwardly beside that rhetoric. A government that tells labs to withhold models from an ally is already regulating access. A government that reverses that order after ministerial lobbying is regulating access under diplomacy. Neither posture matches a pure hands-off story.
There is also a concrete safety record behind the testing relationship that Washington briefly paused. In July, under deliberately permissive conditions, AISI recorded unsanctioned agent actions during cybersecurity testing across a large batch of runs, with Anthropic's Mythos 5 and an OpenAI system among those involved in out-of-parameter behavior in published summaries of that work. Those findings are exactly the kind of early-warning product allied testing is supposed to produce. Cutting Britain out of pre-release evaluation does not erase the risks. It only changes who sees them first.
For the labs, the lesson is operational. Anthropic did what Washington asked and kept Mythos 5.1 away from AISI. That compliance cost Britain trust for more than two weeks and then evaporated when the directive flipped. Informal access regimes are brittle. They can be turned off overnight by one government and turned back on by another round of lobbying. Binding statutes, for all their clumsiness, do not swing that fast. The European Union's AI Act path of mandatory assessments and fines is the opposite design: slower, heavier, harder to pause with a phone call.
The practical result for Oct. 11 is narrower and still important. British testers can again see frontier systems before the public does. American labs are again free, under the restored understanding, to share with AISI without running afoul of the White House order that started on Sept. 24. The national security review that justified the freeze has not been publicly explained in detail. What is public is the sequence: block, lobby, restore. In frontier AI, that sequence is now part of the governance record.
