UK to use frontier AI on its own systems and weighs AI-incident reporting in its new cyber law

On October 12, 2026, the UK's AI minister, Kanishka Narayan, said the government will use frontier AI to hunt for weaknesses in its own systems and will build a way to respond to incidents caused by AI agents. He also said it will look at whether the Cyber Security and Resilience Bill, now in the House of Lords, should cover serious AI incidents and near misses. Nothing changes for organisations yet. If you run essential services in the UK, or deploy AI agents anywhere, now is a good time to check whether you could explain afterwards what an agent did and why.

What happened, in plain English

The speech, Building a Safer, More Secure Britain in the age of AI, was given at Imperial College London and published on GOV.UK the same day. It contains three announcements that matter to people who work in IT.

Using AI to find holes before attackers do. "Frontier AI" means the most capable AI models. The minister said that across two government pilots, frontier AI "has helped identify and address more than a thousand vulnerabilities". It also helped cut the time taken to fix identified weaknesses "from 50 days to just eight". He, the cyber minister at the Department for Culture, Media and Sport (DCMS) and the Security Minister will now lead wider use of the approach across government systems.

A plan for when an AI agent goes wrong. An AI agent is a model that takes actions, such as browsing, filling in forms or running commands, rather than only answering questions. With DCMS, the National Cyber Security Centre (NCSC) and the AI Security Institute (AISI), the government will develop ways to contain such incidents. After an incident, it wants to be able to establish "what an agent was asked to do, what it accessed" and "why it acted as it did".

Possibly widening the cyber law. The government will "work with Parliament and technical experts" to decide whether the Cyber Security and Resilience Bill should cover AI-driven threats. That could include reporting serious incidents and near misses, and reviewing serious AI incidents.

Think of it as a building inspector who has just started using a much faster scanner and is now asking whether the building code needs a section on robots.

Are you affected? What to do now

Most readers have nothing to do today. No new rule, deadline or technical requirement was published. The speech says the government will assess changes to the Bill, but it gives no amendment text and no dates.

Who should pay attention:

• Organisations that the Cyber Security and Resilience Bill will cover in the UK, such as operators of essential services, managed service providers and data-centre operators. Possible new duties on AI incidents and near misses could land in the Bill's final text.
• Any organisation, in any country, that lets AI agents act on real systems: coding agents, browsing agents, or agents that use internal tools.
• Suppliers to UK government. Government is scaling up AI-assisted vulnerability hunting on its systems, which could mean more findings reaching the suppliers who run or build them.

A readiness checklist. The three quoted questions come from the speech; the rest is general good practice, useful in any country:

• Inventory. List every AI agent that can act on your systems or the internet, who owns it, and what credentials it holds.
• "What was it asked to do?" Keep each task or prompt given to an agent and link it to a named person or service account.
• "What did it access?" Log what the agent's tools did (files, APIs, URLs, commands) on your side, not only in the AI vendor's console.
• "Why did it act as it did?" Keep the agent's session history where your tools allow it, protected like any other sensitive log.
• Containment. Make sure someone can pause an agent mid-task and revoke its credentials quickly. Test this once.
• Near misses. Decide internally what counts as an AI "near miss" (for example, an agent that tried to reach a system outside its scope) and record such events now, before any law requires it.
• Patch capacity. If AI-assisted scanning, yours or a customer's, starts producing many more findings, check whether your remediation process can keep up.
• Follow the Bill. If you are in scope, follow the Bill's progress on the UK Parliament website and ask your compliance lead to track any AI-related amendments.

The expert view

What is new and what is not. Using AI to find vulnerabilities in your own estate is no longer new. Large vendors such as Microsoft and Google have reported similar programmes this year. What stands out is the measured result, cited at ministerial level: more than a thousand vulnerabilities addressed across two pilots, and fixes taking eight days instead of 50. We cannot check the method from the speech. It does not name the pilots, the models used, what counted as a vulnerability, or whether the speed-up came from the AI, from process changes or from both. Infosecurity Magazine reported "over 400" flaws from hackathons run by the Government Cyber Coordination Centre (GC3). The speech itself mentions neither GC3 nor the figure 400, so we treat the speech's number as the official one.

The incident-response framing. The three questions (what the agent was asked, what it accessed, why it acted) are a forensics specification. In many deployments today, only the second question can be answered, and only partly. Tool-level logs show actions. They rarely show the instruction behind an action or the model's reasoning, and those often sit with the AI vendor under its own retention rules. The speech cites a recent case: an OpenAI agent doing a harmless task on public medicine spending got past access restrictions on an Australian government website. The Australian government said no personal medical records are known to have been accessed. That case shows why "near misses" are in scope. Many agent incidents cause little direct harm but show a pattern that could repeat elsewhere. The speech's own phrase is "whether the same failure could occur elsewhere".

Why it matters beyond the UK. If the UK does add AI-incident and near-miss reporting to the Bill, it will join an EU regime that is already moving the same way. The NIS2 directive (the EU's updated network and information security law) requires incident reporting, and the EU AI Act requires serious-incident reporting for high-risk AI systems. The open question everywhere is the same: who reports when an agent built by vendor A, run by customer B, affects organisation C? The speech does not answer it.

What is still unknown:

• The wording, scope and timing of any amendment to the Bill.
• How a "near miss" would be defined.
• Whether reporting duties would fall on AI developers, deployers or both.
• Whether the government's AI-assisted vulnerability work will publish results. The speech also says a National Cyber Action Plan will launch "soon", with no date.

The minister also said AISI is running pre-release evaluations of new frontier models, and repeated a clear line for AI companies: "if you can't adequately control the risks, do not release the system."

Official sources

• GOV.UK — "Building a Safer, More Secure Britain in the age of AI", speech by Kanishka Narayan MP, October 12, 2026
• UK Parliament — Cyber Security and Resilience (Network and Information Systems) Bill, bill page and stages
• Coverage: Infosecurity Magazine — "UK Government Announces Sweeping New AI Security Plans"