On October 11 the White House restored pre-release frontier models to Britain's AI Security Institute, ending a 17-day freeze that began on September 24 (AI Weekly). OpenAI, Anthropic, and Google DeepMind are on the pre-release list again.
That list returned because AI minister Kanishka Narayan lobbied Washington for the reversal. He says those conversations leave him confident the institute can test the latest generation. He also says dependence on another country's technology means you shouldn't treat the restored access as permanent (Quartz).
What the hold covered
On September 24 the Trump administration asked OpenAI and Anthropic to stop sharing new frontier models with the institute until a US national security review was finished. A senior US official said the pause matched policy for every new frontier model, because the companies are American.
Google DeepMind wasn't named in that September request. It still isn't clear whether the review finished, or what it found (CryptoBriefing). Britain stood up the institute after the 2023 AI Safety Summit so public researchers could test advanced systems, and the risks they carry, before a release. Before the ask, British testers saw new models under a voluntary arrangement, early enough to probe a system before the public did. The hold broke that habit.
The hold ended on October 11, and the list that returned is wider than the one that was paused.
Claude Mythos stayed out
Anthropic didn't share Claude Mythos 5.1 with the institute while the hold was in force. Access stayed with a set of US organizations, and the company said it was coordinating with the US government to widen that circle.
Institute director Henry de Zoete acknowledged that AISI hadn't received Anthropic's latest model during the freeze. He also said the institute had tested OpenAI's GPT-6 Astra before release. The Mythos gap was a cutoff on one new model, after the institute had already tested another lab's system early. Whether that wider circle now includes the institute is still unsettled.
The pilots and the bill
Scientists at the institute are now conducting rigorous evaluations of the latest generation of advanced AI models, following work with leading AI developers and the US government (GOV.UK). Narayan said that at Imperial College London's Base One incubator. In his first speech since the AI job was raised to Cabinet level, he named Google's Gemini (The Independent).
Google had said it would limit Gemini to a vetted group of cyber specialists, so hackers couldn't misuse it. Narayan said UK scientists would examine that model anyway.
Alongside the access line, he set out what Britain wants to do with the models it can see. Two pilots that put frontier AI on government systems have already identified and addressed more than 1,000 vulnerabilities. Narayan said he would work with other ministries to find weak spots in government systems before an attacker does. Extending it is meant to protect public services people use every day, and to support longer-term adoption across the public sector.
Ministers are also looking at the Cyber Security and Resilience Bill for ways to identify and report serious AI incidents and near-misses. That includes loss of control, where an autonomous system slips human oversight. Work with the National Cyber Security Centre and the institute is aimed at incidents involving AI agents, which can spread faster and wider than the emergencies governments already train for. Any reporting rules still have to be settled with parliament, technical experts, and the operators who would carry them out.
He told the companies building these systems, "If you cannot adequately control the risks, do not release the system."
If you're waiting on the next pre-release handoff from Anthropic or OpenAI, watch that handoff. Another freeze would mean the posture that opened this one is the default. Treat allied testing as something Washington can grant and take back. Don't plan a product, a procurement, or a safety review as if early UK access renews on its own.
Photo: Eden, Janine and Jim / Wikimedia Commons, CC BY 2.0.
