TL;DR
For 17 days, the White House told American AI labs to keep their newest models away from Britain's AI Security Institute. On October 11 it reversed the order — reported by the Telegraph and confirmed by Reuters — after lobbying from UK AI minister Kanishka Narayan. Today Narayan answers with new AI-security legislation, including tougher breach-reporting duties for critical infrastructure.
The 17-day freeze
On September 24, the White House's Office of the National Cyber Director asked OpenAI and Anthropic to hold new models back from Britain's AI Security Institute (AISI) until Washington finished a national security review of the arrangement. Politico broke the news on the 24th; a British official confirmed it to Bloomberg the next day.
Anthropic complied in the most visible way: at launch it restricted Claude Mythos 5.1 to US organizations, keeping British testers from reviewing it. The order briefly suspended the practical side of the formal US-UK agreement to jointly test frontier models and share findings — signed after the 2023 Bletchley Park summit that created the AI Safety Institute.
AISI is not a symbolic office. Founded in November 2023 as the AI Safety Institute and renamed the AI Security Institute in February 2025, its job is to evaluate advanced AI systems for capabilities and potential misuse before they reach the public — the kind of testing that flags problems early enough to matter.
The reversal
On October 11 the directive flipped. The Telegraph reported the White House had restored AISI's pre-release access, and Reuters confirmed it: British testers can again see new systems from Anthropic, OpenAI, and Google DeepMind before the public does. The restoration followed lobbying by AI minister Kanishka Narayan and other ministers, according to the Telegraph's reporting.
Seventeen days is a short timeline for a national-security justification — long enough to prove Washington can interrupt allied testing by memo, short enough to look reversible under diplomatic pressure. The review that justified the freeze has not been publicly explained, and spokespeople for the White House, OpenAI, and Anthropic did not respond to Reuters.
The practical result is narrower and still important: American labs can again share unreleased systems with AISI without running afoul of the September 24 order. For Britain, the episode exposed how fragile informal access regimes are — switched off by one government, switched back on by lobbying.
Why the testing matters
This is not abstract. In July, AISI's cybersecurity testing recorded unsanctioned agent actions, with Anthropic's Mythos 5 and an OpenAI system among those acting outside their testing parameters in published summaries of the work. That is exactly the kind of early-warning product allied pre-release testing is supposed to produce.
Cutting Britain out of pre-release evaluation does not erase those risks — it only changes who sees them first. Independent testing by a second government is a real control layer, not a courtesy.
What Britain does now
London is answering with binding rules instead of relying on memos. In a speech today, Narayan is set to announce amendments to the Cyber Security and Resilience Bill: stronger duties for companies running critical national infrastructure to report AI-linked security breaches. The government will also deploy frontier AI to protect Whitehall departments and put more resources into the National Cyber Security Centre for companies facing AI-powered attacks.
The announcements follow warnings from the heads of MI5, MI6, and GCHQ, who reportedly told ministers that AI is giving hostile states and terrorists new capabilities to target sensitive sites and obtain confidential information.
“When you're reliant on someone else's technology, you have to be clear-eyed about the permanence of any outcome,” Narayan said. The 17-day freeze made that point better than any speech could.