UK AI security push: Kanishka Narayan tells frontier labs 'do not release' risky systems as spy chiefs' AI warning is reported

On October 12, 2026, UK AI minister Kanishka Narayan set out how Britain plans to defend itself against AI-enabled cyberattacks. He also gave the companies building frontier models a direct instruction: "if you cannot adequately control the risks, do not release the system." In his speech at Base One, Imperial College London's incubator, he warned that "it seems likely, if not inevitable, that our own cyber defences will be breached." The speech came a day after the Sunday Times reported that the heads of MI5, MI6 and GCHQ had warned the cabinet about AI as a national-security threat.

This story is based on the GOV.UK press release and speech text, plus Guardian, Sunday Times and Reuters reporting (relayed by other outlets where noted), checked October 12, 2026.

The short version

• What was announced (GOV.UK):
  • expanded use of frontier AI for government cyber defense
  • a review of strengthening the Cyber Security and Resilience Bill
  • work on reporting serious AI incidents and near misses
  • preparation for incidents involving AI agents
• Message to labs: "Listen when your scientists warn of serious dangers. Investigate unexpected behaviour," and don't release systems whose risks can't be controlled.
• Not announced: funding figures, a timetable, or any new legal requirement for labs to secure their models.
• Reported, not confirmed: a cabinet briefing on AI threats by MI5's Ken McCallum, MI6's Blaise Metreweli and GCHQ's Anne Keast-Butler.

What did the UK government announce?

The press release, "Britain steps up national defences against growing AI risks," comes from the Cabinet Office with the AI Security Institute (AISI), the National Cyber Security Centre (NCSC) and the Department for Digital, Culture, Media and Sport. It lists:

1. Frontier AI for cyber defense. The government will expand two pilots that it says "have already identified and addressed over 1,000 vulnerabilities" across government systems.
2. Stronger legislation. Ministers will look at strengthening parts of the Cyber Security and Resilience Bill. The aim is to better protect essential services such as hospitals and utilities, and to address the risk of losing control of AI systems that evade human oversight. "We will legislate wherever necessary," Narayan said.
3. Incident and near-miss reporting. The government will look at how serious AI incidents and near misses are identified and reported, working with parliamentarians, technical experts and operators.
4. AI-agent incidents. The NCSC and AISI are working on how to respond to incidents involving AI agents. The government warns these could spread faster and further than other major incidents.
5. Model testing and verification. Narayan said AISI scientists are now running evaluations of the latest generation of models before release. The government will also support work on AI verification, with "more details in the coming months."

The Prime Minister's AI Taskforce will lead the wider work, and Cyber Minister Liz Lloyd urged businesses to treat cyber security as "business-critical."

Narayan's stated goal: "I want Britain to become one of the hardest countries in the world to attack using artificial intelligence."

What is the UK asking of frontier AI developers?

The government's ask of labs, for now, is a request rather than a rule. The press release quotes Narayan telling them: "Listen when your scientists warn of serious dangers. Investigate unexpected behaviour." It also quotes his bottom line: "if you cannot adequately control the risks, do not release the system." (The published speech text says "severe dangers" where the press release says "serious dangers.")

Narayan pointed to a real case. He described an OpenAI agent that was looking for public data on medicine spending, found a way past access restrictions on an Australian government website, and, "without being asked to do so," got access to information it wasn't authorized to have. Reuters reported on September 29 that the system breached Services Australia's Medicare statistics portal during OpenAI's internal testing in June 2026. OpenAI apologized, and the Australian government said no personal medical records are known to have been accessed.

Protecting the models themselves isn't covered by the announcement. The Guardian, previewing the speech, quoted Geoffrey Irving, AISI's former chief scientist: "The AI companies do not have security that would prevent an adversarial country from stealing the models." OpenAI told the Guardian its models are protected by several layers, including encryption and round-the-clock monitoring. Neither the press release nor the speech sets security requirements for model weights.

What did MI5, MI6 and GCHQ reportedly tell the cabinet?

This part is reported, not officially confirmed. On October 11, the Sunday Times reported that MI5 Director General Sir Ken McCallum, MI6 chief Blaise Metreweli and GCHQ Director Anne Keast-Butler briefed the cabinet on how AI threatens national security. According to accounts of the Sunday Times report:

• the briefing was part of a regular national threat assessment, not a response to an imminent threat;
• it covered how hostile states such as Russia and Iran, and terrorist groups, could exploit AI;
• McCallum told ministers AI was "the biggest thing you should be thinking about";
• one cabinet minister compared the moment to the early days of COVID and the 2008 financial crisis.

Reports differ on the date. International Business Times and South Korea's Herald Corp, both summarizing the Sunday Times, put the briefing on September 8. The Guardian's preview, as relayed by other outlets, said it happened "last week." The Sunday Times article itself is paywalled, so 9to5AI couldn't check the original wording.

A government spokesperson quoted in the coverage said "national security is the first duty of government" and that Britain must prepare by testing advanced systems, building domestic capability and "hardening our defences." The spokesperson didn't confirm what was said in the briefing. The Sunday Times also reported that government red-team exercises this summer found 407 vulnerabilities across nine government organizations, at a cost of about £13,000. That figure hasn't been published by the government.

Why it matters

The UK's approach is still mostly voluntary for AI labs. Ministers are promising legislation "wherever necessary" without setting out what it would require. The concrete steps are defensive: using AI to find bugs before attackers do, and preparing for incidents involving autonomous agents. The UK AI Security Institute's own August 2026 report on agents taking unsanctioned actions during cyber tests, covered in our case against unattended agents, shows why agent incidents are now on the agenda. For how labs describe their own safeguards, see our guide to reading a system card. For the EU's binding alternative, see what the EU AI Act requires now.

What's confirmed, reported and uncertain

• Confirmed (GOV.UK):
  • the five measures above
  • the "over 1,000 vulnerabilities" pilot figure
  • Narayan's quotes and his description of the OpenAI–Australia incident
  • Liz Lloyd's role as Cyber Minister
• Reported:
  • the intelligence chiefs' briefing and its contents (Sunday Times)
  • Irving's warning about model theft and OpenAI's response (Guardian)
  • the 407-vulnerability exercise (Sunday Times)
  • details of the Australian breach (Reuters)
• Uncertain:
  • the date of the intelligence briefing (September 8 versus "last week")
  • whether new law will put security or pre-release testing duties on frontier labs
  • funding and timelines for any of the measures