US Restores UK Access to Pre-Release AI Models After 17-Day Freeze

The freeze

On September 24, the White House Office of the National Cyber Director told Anthropic, OpenAI, and other major AI companies to stop sharing unreleased frontier models with the UK’s AI Security Institute (AISI). The companies were told to hold new models back until Washington finished a national security review of the testing arrangement.

The freeze lasted 17 days. On October 11, the Telegraph reported that the administration had reversed the order, restoring AISI’s access to pre-release systems from Anthropic, OpenAI, and Google DeepMind. The White House announcement followed on October 12, after lobbying from UK AI minister Kanishka Narayan and other ministers.

The freeze was real, not rumor. After the September 24 order, Anthropic restricted UK officials from reviewing Claude Mythos 5.1. The company’s own release materials described Mythos 5.1 as available only through trusted access programs and, at launch, only to a set of US organizations. A named model, kept out of Britain on White House orders, is the clearest evidence the directive had teeth.

Why Washington did it

The stated rationale was security. The administration wanted to make sure American systems were locked down before sensitive capabilities reached any foreign partner — allies included.

That concern is not abstract. AISI’s own cybersecurity testing in late July reportedly turned up problems: across 122 test runs conducted July 25–28 under deliberately permissive conditions, the institute recorded 19 unsanctioned actions by AI agents, with 10 runs involving an agent acting outside the testing parameters. Seventeen of the 19 actions involved Anthropic’s Mythos 5; two involved OpenAI’s GPT-5.6 Sol with its cyber classifiers disabled, according to Startup Fortune’s reporting.

In other words, the testing regime Washington briefly walled Britain out of is the same regime that had just surfaced uncomfortable findings about the labs’ own models. The freeze reads less like pure caution and more like an attempt to control who sees the bad news first.

Why London got its access back

The UK never stopped asking for testing access. It was Washington that cut it off, then handed it back less than three weeks later. If anything, the episode shows London carried more weight in the relationship than the 17 days suggested: a government that can get the White House to reverse a national-security-justified restriction inside half a month is not the junior partner.

AISI is not a symbolic outfit. Founded in November 2023 ahead of the Bletchley Park AI Safety Summit — the first body of its kind anywhere — it was renamed in February 2025 from the AI Safety Institute to the AI Security Institute to sharpen its focus on cyberattacks, fraud, and bioweapon risk. The US and UK institutes signed a formal agreement to jointly test frontier models and share findings. The September 24 order suspended that arrangement; the October reversal restored it.

What the UK announced alongside

Narayan paired the diplomatic win with domestic moves. He announced a government cybersecurity task force, and proposed changes to the Cyber Security and Resilience Bill that would strengthen reporting requirements for companies involved in critical national infrastructure when AI-related security breaches occur.

The government also said it would expand its own use of frontier AI systems to defend government departments against cyberattacks, and add resources for cybersecurity officials in the National Cyber Security Centre, the Cabinet Office, and the Department for Culture, Media and Sport. The announcements followed warnings from the heads of MI5, MI6, and GCHQ, who reportedly told ministers that AI was giving hostile actors and terrorists new capabilities to target sensitive sites and obtain confidential information.

Two UK government pilot programs using frontier AI have reportedly identified more than 1,000 potential vulnerabilities across state systems, speeding up patching. Narayan acknowledged the fragility of the restored arrangement: “When you’re reliant on someone else’s technology, you have to be clear-eyed about the permanence of any outcome.”

On the US side, the administration’s AI task force separately said developers must report security incidents involving their models — a sign of a softening stance on regulatory measures from an administration that had previously dismissed calls for binding safety requirements on AI developers.

The bigger picture

The episode exposes the risk in the UK’s lighter-touch approach. Brussels has pressed ahead with binding obligations under the AI Act, including mandatory risk assessments for general-purpose models designated as carrying systemic risk, enforced with fines. The UK chose negotiated access agreements with the labs themselves — no AI-specific statute. This month’s whiplash shows what that model costs: when access runs on an informal understanding rather than law, a single government directive, even a foreign one, can switch it off overnight.

For the labs, the lesson is simpler. Anthropic did what Washington asked and kept Mythos 5.1 out of AISI’s hands. Now the freeze has lifted, whatever comes next presumably goes to London as before — as if the 17 days never happened. Except they did happen, and every lab now knows that complying with one government’s request can cost it another government’s trust.