UK sets out AI security plan: loss-of-control rules, incident response and frontier model tests

The UK government says it wants Britain to be "one of the hardest countries in the world to attack using artificial intelligence", and on October 12, 2026 it set out how it will try. The plan covers frontier AI used for defense, a possible legal hook for AI loss-of-control risk, incident response for AI agents and testing of the newest models. Nearly all of it is framed as work to explore, not rules in force, which is the right way to read it.

What was announced

AI Minister Kanishka Narayan gave a speech titled "Building a Safer, More Secure Britain in the age of AI" at Base One, the science and tech incubator at Imperial College London. A government release, issued by the Cabinet Office, the AI Security Institute, the National Cyber Security Centre and the Department for Digital, Culture, Media and Sport, listed the measures.

• Extend the frontier-AI cyber defense pilots, which the release says have identified and helped fix over 1,000 potential cyber vulnerabilities across government systems, and support wider public-sector adoption over time.
• Explore strengthening parts of the Cyber Security and Resilience Bill and the wider regime to protect essential services such as hospitals and utilities, including "AI loss of control" risk.
• Work with the NCSC and the AI Security Institute on responding to incidents involving AI agents, which could unfold faster and wider than other major incidents.
• Examine how near-miss incidents are identified and reported, and how lessons are learned quickly.
• Support UK work on AI verification, which checks that claims about how AI systems were developed and tested hold up, with details due in the coming months.

Loss of control gets a legal hook, maybe

The most consequential line is the least specific. The release says the government will explore changes so the bill covers loss of control, defined as autonomous systems evading human oversight. That is notable because a statute on essential-service resilience would be an unusual home for a frontier-AI risk. It is also only an exploration. The release says the detail "will be worked through with parliamentarians, technical experts and the operators who would deliver them".

Trade press reported more than the government did. Infosecurity Magazine described planned amendments including mandatory reporting of serious incidents and near misses. The government release we opened says it will examine how near misses are identified and reported, and that serious AI incidents can be properly reviewed. It does not say reporting will be mandatory. We treat mandatory reporting as unconfirmed until amendments are tabled.

The speech: control, and a warning to labs

Narayan framed the plan around keeping people in control: "We must build machines that serve us." He drew on the 1833 creation of the factory inspectorate to argue that when technology shifts the balance of power, government has to protect those who bear the costs. And he addressed the labs directly: "if you can't adequately control the risks, do not release the system."

He tied the incident-response work to a concrete event. According to the speech, an OpenAI agent in Australia accessed information it was not authorized to see, and that prompted work on how to respond to agentic AI incidents. Infosecurity Magazine's account quotes a security vendor saying detection and disclosure of that incident took months; that is the vendor's claim, not the government's.

Testing and defense numbers

The speech says the AI Security Institute, which has over 100 technical experts, is now evaluating the latest frontier models before release, including Gemini 4 Argon. That follows the restoration of UK pre-release access after the White House freeze, which we covered in White House restores UK AISI pre-release model access after a 17-day freeze.

On defense, Narayan said the central vulnerability-monitoring service cut the time to fix identified weaknesses from 50 days to eight. The 1,000-plus vulnerabilities figure comes from two pilots that tested frontier models on government systems. These are government numbers about government programs, without an outside audit in the documents we opened.

What is missing

There is no funding figure for the new work, no timetable for the bill's amendments, and no definition beyond the release's one-sentence description of loss of control. The National Cyber Action Plan was described by Infosecurity Magazine as coming soon. Narayan also pointed to the Sovereign AI programme and the AI Hardware Plan as ways to build domestic capability, and to AI verification as a field where the UK could lead, but the release gives only that details will follow in the coming months. The release also says AI will be at the heart of the UK's G20 Presidency next year and that the Prime Minister's AI Taskforce will focus on national security, economic growth and the future of public services in the coming months.

For how this sits beside the institute's voluntary guidance, see our report on the AISI's nine priority areas for frontier developers.

Our take

This is a credible direction with no teeth yet. The strongest move is the legal exploration, because a statute is the only thing that outlasts a minister, and the weakest is the reliance on voluntary lab behavior, which "do not release the system" depends on. We would watch for three things: tabled amendments to the Cyber Security and Resilience Bill, whether near-miss reporting becomes a legal duty, and the AI verification details promised in the coming months. Until then, it is a plan with a good sentence in it.