Summary
On October 12, 2026, UK AI Minister Kanishka Narayan confirmed that scientists at the UK AI Security Institute (AISI) were again evaluating the latest generation of frontier models before release, "following work with leading AI developers and the US Government." Bloomberg had reported on September 25 that the Trump administration told OpenAI and Anthropic to withhold new models from the institute, and on October 11 the Telegraph reported that the administration had reversed that request. The confirmation came in a UK government package that also puts AI incident reporting, AI loss-of-control risk and agent incident response on the agenda for the Cyber Security and Resilience Bill regime.
What Happened
AISI has received pre-deployment access to frontier models from US laboratories under voluntary arrangements since its founding after the 2023 AI Safety Summit. On September 25, 2026, Bloomberg reported that the Trump administration had told OpenAI and Anthropic to withhold new models from the UK agency. OpenAI declined to comment to Bloomberg, and the White House did not respond.
On Sunday, October 11, the Telegraph reported that the administration had reversed the request and that US developers could again give AISI early access to models for safety testing.
On Monday, October 12, the UK government published a press release under the names of the Cabinet Office, AISI, the National Cyber Security Centre and the Department for Digital, Culture, Media and Sport to accompany a speech by Narayan at Imperial College London's Base One incubator. The release states that "scientists at Britain's AI Security Institute are now conducting rigorous evaluations of the latest generation of advanced AI models, following work with leading AI developers and the US Government." Bloomberg reported the announcement as the UK regaining the ability to evaluate leading models before release, reversing the US policy. TechCrunch, reporting the same day, wrote that AISI "confirmed on Monday that it is able to access and test the latest models."
The release paired the access confirmation with a set of domestic measures. The government said it would extend two pilots that used frontier AI to identify and fix more than 1,000 potential vulnerabilities in government systems. It said it would explore strengthening parts of the Cyber Security and Resilience Bill to cover emerging AI risk to essential services such as hospitals and utilities, explicitly including "AI loss of control risk, where autonomous AI systems evade human oversight," and would look at how serious AI incidents and near misses can be identified and reported. The National Cyber Security Centre and AISI were tasked with working out how incidents involving AI agents should be responded to, and the government said more detail on support for AI verification would follow "in the coming months."
Narayan's statement included a message to developers: "Listen when your scientists warn of serious dangers. Investigate unexpected behaviour. And if you cannot adequately control the risks, do not release the system." He also said: "I want Britain to become one of the hardest countries in the world to attack using artificial intelligence."
Why It Matters
The hold extended US control over who may see frontier models before release from foreign users to an allied government's testing body. In June 2026 the Commerce Department ordered Anthropic to suspend Fable 5 and Mythos 5 access for all foreign nationals, and the UK's request for an exemption for British nationals was declined at the G7. The September request targeted pre-release evaluation by AISI itself, the arrangement the UK has built its frontier-safety policy around since 2023. AISI's access rests on voluntary company agreements, so the reversal restores the previous arrangement without creating any legal right of access, and a future administration request could interrupt it again.
The UK confirmation does not name the models under evaluation, and no US government statement on either the original request or its reversal has been published; the account of both rests on press reporting and the UK government's description. Which models AISI was unable to test during the hold, and whether any were released without UK evaluation in that period, has not been officially disclosed.
The accompanying measures place AI loss of control and agent incidents inside the UK's existing cyber legislation rather than a standalone AI bill. They arrived three days after the White House told AI companies that they must disclose security incidents with their models, so incident reporting now features in both governments' stated policy, while the access arrangement between them remains informal.
§ How to read the metadata
- Landmark
- Fundamentally alters the trajectory; 2–5 per year.
- Major
- Meaningfully shifts the landscape; 2–4 per month.
- Notable
- Worth documenting; significance can be upgraded later.
- Confidence
- High = primary sources corroborate. Medium = credible secondary only. Low = provisional. Disputed = credible sources disagree.
- Contestation
- Uncontested = no formal challenge. Contested = at least one challenge open. Superseded = replaced by a later entry. Unresolved = dispute still open.