Britain's AI Security Institute is evaluating the latest generation of frontier models again, including Google's Gemini 4 Argon. AI Minister Kanishka Narayan confirmed the evaluations on Monday, October 12, in a speech at Imperial College London's Base One incubator, as the government unveiled a wider package to put frontier AI to work on national cyber defences and to prepare for serious incidents involving autonomous agents. The government said the work follows cooperation with leading developers and the United States government, and that the institute's scientists are testing for risks early and probing whether safeguards actually hold.
That confirmation matters because the access had become uncertain. Reuters reported that last month the White House asked OpenAI and Anthropic to withhold new models from British testers pending a United States review, and in a September letter the institute noted that Anthropic had made clear no organisation outside the United States had access to its Mythos 5.1 model. Narayan was candid that a country reliant on someone else's technology has to be clear-eyed about how permanent any access arrangement is. In other words, pre-release testing is no longer just a technical courtesy between labs and evaluators. It is now negotiated statecraft, and it can be switched off.
The defensive case for keeping that access is concrete, not theoretical. The government said two pilots using frontier AI have already helped identify and fix more than a thousand potential cyber vulnerabilities across government systems and detect potential threats, and it now wants to extend that programme to protect the public services people rely on every day. Reporting on the minister's speech added that a central vulnerability-monitoring service has cut the time taken to fix identified weaknesses from 50 days to eight. That is the kind of operational gain builders should pay attention to: the same models that create new attack surface are also, in practised hands, shrinking patch windows dramatically.
The rest of the package points toward harder rules. Ministers will explore strengthening the Cyber Security and Resilience Bill regime so essential services such as hospitals and utilities are better protected from AI-related risks, including loss-of-control scenarios where autonomous systems evade human oversight. Work is also underway with the National Cyber Security Centre and the AI Security Institute on how incidents involving AI agents should be responded to, how near misses can be identified and reported, and how lessons can be learned quickly. Narayan's message to labs was blunt: listen when scientists warn of serious dangers, investigate unexpected behaviour, and do not release a system whose risks cannot be adequately controlled. The government also flagged AI verification, checking whether claims about how a system was developed, tested and assessed stand up to scrutiny, as a field Britain wants to lead, with more detail promised in the coming months. Microsoft President Brad Smith, in a BBC interview published the same day, said Britain was well placed to help lead work on AI safety guidelines.
For anyone building with AI, the practical lesson is to design for outside scrutiny from the start. Expect pre-release evaluations, incident and near-miss reporting, and verification of safety claims to become normal gates, not exceptional requests, especially for agentic systems that act with little human supervision. Keep tamper-evident logs, document evaluations honestly, and know in advance how you would pause or roll back an agent that behaves unexpectedly. Teams that treat that evidence as a product feature will move faster through the new gates. Teams that treat it as paperwork will discover, as the British testers just did, how quickly access and trust can disappear.
