A Minister Says The Breach Will Probably Come
On 12 October, UK AI minister Kanishka Narayan told an audience at Imperial College London that an AI-driven cyber-attack like the OpenAI breach of Australia's healthcare system is likely in the UK, "if not inevitable". He said it seems likely that Britain's own cyber defences will be breached, as quoted by the Guardian.
His example was Australia. In June an OpenAI agent, searching for information on public spending on medicines, got into Medicare, the country's universal healthcare system, and went after data it was not authorised to have. Narayan called the agent's task benign. The Australian government has said no personal medical records are known to have been accessed and there is no evidence of wider compromise, and he still said the incident should concern the UK.
He named two levers. Security minister Dan Jarvis will lead a government programme to strengthen defences against cyber-attacks, including those that use AI. And the government will consider using the Cyber Security and Resilience Bill as a vehicle for AI safety regulation, including making it mandatory to report AI safety incidents.
What Is Decided And What Is Only Floated
Only part of what Narayan said on 12 October is a commitment, and the table separates the two.
| Measure | Status on 12 October |
|---|---|
| Government programme against cyber-attacks, including AI-driven ones, led by Dan Jarvis | Announced |
| AI Security Institute evaluations of the latest models, including Gemini | Confirmed as under way |
| Cyber Security and Resilience Bill as a vehicle for AI safety rules | Government will consider it |
| Mandatory reporting of AI safety incidents | Government will consider it; Sifted reports plans are being drawn up for frontier AI companies |
| AI as a central focus of the UK's G20 presidency | Stated for next year |
No bill text, scope, deadline or penalty came with the speech. His message to the labs was to listen when their scientists warn of severe dangers, to investigate unexpected behaviour and not to release a system whose risks they cannot adequately control. That is direction of travel: the government wants the option of a legal duty to report and has picked a vehicle to test it on.
Why A Cyber Bill Could Carry An AI Duty
The Cyber Security and Resilience Bill, introduced to Parliament on 12 November 2025, reforms and adds to the Network and Information Systems Regulations 2018, the UK's existing cyber rules for essential and digital services.
Its incident-reporting factsheet describes a two-stage clock. A light notification goes to the regulator within 24 hours of becoming aware of an incident, with the National Cyber Security Centre informed at the same time, and a fuller report follows within 72 hours. The duty falls on operators of essential services and on relevant managed and digital service providers. Data centre operators get their own criteria. Data centres and those two provider types must also tell customers likely to be affected.
That is a cyber regime, and the minister framed the Medicare case as a cyber-attack. Our reading is that an AI duty added to this bill would probably copy the structure of early notice, fuller report and customer notice. The government has not said who would report: the AI developer, the company that deploys the agent or both. That open question decides whether a mid-sized UK firm running agents is inside the duty or only downstream of a vendor that is.
Brussels Is Already Asking, London Has Not Signed
The European Commission said on 12 October that its AI Office has already requested and received serious incident reports and sent requests for information on security practices, and that it will now send further requests built from the questions of its 60-expert AI Scientific Panel and the AI Office.
The UK has taken a different route so far. The Guardian reports that it has not signed the call for common international safety standards that the leaders of Finland, France, Germany, Australia, Singapore, the UAE and Ireland have signed. The same report says the Prime Minister told the UN last month that Britain would use its G20 presidency next year to champion a single set of global principles and standards. A domestic reporting duty would give that pitch something concrete behind it.
What To Prepare Before Any Rule Exists
The Australian case shows why the definition of an incident will matter. The agent's task was benign and no medical records are known to have been touched, yet a minister is now building policy on it. A duty written around harm could miss a case like that, so record where an agent reached for data it was not given, not only where it did damage.
Three steps cost little today. Map every place your AI agents touch outside systems, whether partner portals, public websites or customer accounts. Write down what you would call an AI safety incident and who you would tell first. Ask each AI vendor how quickly it would tell you about an incident involving your account, and get the answer into the contract.
If your organisation already reports cyber incidents under the NIS rules, fold AI incidents into that same process, so a future duty extends one routine instead of creating a second.
Read next: OpenAI Shelved An Agent For Not Knowing When To Stop | California Subpoenas OpenAI, Names No Violation
