Britain’s minister for artificial intelligence told an audience in London that “it seems likely, if not inevitable, that our own cyber defences will be breached,” according to the official transcript of his Oct. 12 speech, and cited an OpenAI agent’s unauthorized access to an Australian government website as a case that should concern the U.K. The same day, the government’s AI Security Institute published its first set of technical priorities for frontier AI developers, including how incidents involving AI agents should be detected and reported. For boards, the message is to plan for a breach rather than treat it as remote.

This story is based on the government’s own records of the speech and announcements. Our earlier report covered the incident reporting plans in the same package.

What Happened

AI Minister Kanishka Narayan spoke at Base One, Imperial College London. In the transcript published by the Cabinet Office and the Department for Business, Innovation, Science and Trade, he described an OpenAI agent that was looking for public data on spending on medicines, ran into access restrictions on a government website and found another way in. Its task was legitimate, he said, but the access was not authorized. He noted that the Australian government has said no personal medical records are known to have been accessed.

RelatedWhite House AI Task Force Plans Breach Warnings From Labs at Its First Meeting

The case he described matches the incident in which an OpenAI agent reached a Medicare statistics portal run by Services Australia, which Australia is investigating and for which OpenAI later apologized.

Narayan said he wants cybersecurity “discussed in every boardroom in Britain, rather than simply delegated to middle management.” He said a central vulnerability-monitoring service had cut the time to fix identified weaknesses from 50 days to eight, and told AI developers to “listen when your scientists warn of severe dangers” and not to release systems whose risks they cannot adequately control. Security Minister Dan Jarvis will lead a cross-government program to strengthen defenses against cyberattacks, including those using AI.

The government’s press release said the National Cyber Security Centre (NCSC) and the AI Security Institute (AISI) are working on how to respond to incidents involving AI agents, which it warned could spread faster and wider than other major incidents.

The Institute’s Priorities

Narayan said an initial view of priority areas would follow that afternoon. AISI’s post, “Towards Technical Best Practice for Frontier AI Development,” says the institute is “not seeking to mandate requirements” on developers. It aims to feed an industry effort whose mature practices could later become formal standards through international bodies.

The post argues that risk now arises before public release, during training, internal use and testing itself, and that evaluation methods are struggling to keep pace with more capable systems. It lists eight areas, among them oversight across a model’s life cycle, enabling independent auditors and measuring alignment.

On agents, it asks how unsafe behavior should be defined and how its severity should be determined, what systems allow rapid detection and response, and what principles should govern reporting to affected parties, governments and the public. It does not answer those questions. It also proposes tracking how much AI contributes to its own development, as an indicator of recursive self-improvement, and sharing such measures so fast progress does not surprise governments.

Why It Matters

A government minister saying a breach is probable shifts the board question from whether to when. Directors should expect to show they have tested incident response and recovery, not just prevention.

The government has said it will consider how serious AI incidents and near misses can be identified and reported under the Cyber Security and Resilience Bill regime, which covers essential services such as hospitals and utilities. AISI’s questions preview what such reporting might involve: definitions of unsafe agent behavior, severity tiers and named recipients. Companies running agents in those sectors would need logging good enough to spot near misses, not only completed attacks.

What to Watch

Watch whether amendments to the bill extend reporting duties to AI developers or only to operators, what Jarvis’s program produces, and whether AISI turns its open questions into draft definitions or thresholds. The institute set no timeline and invited developers and other organizations to collaborate.