It’s ‘likely, if not inevitable’, that UK Government cyber defences will be breached by AI, suggested Minister for Artificial Intelligence Kanishka Narayan in a speech at Imperial College London.
He was referring to a breach by an OpenAI agent of an Australian government website.ย He said: “Our National Cyber Security Centre [NCSC] has warned that AI will almost certainly make elements of cyber intrusion more effective and efficient. So my colleague, Dan Jarvis, the [Home Office] Security Minister, will be leading a cross government program to strengthen our defences against cyber attacks, including from AI.ย We published the Government Cyber Action Plan earlier this year addressing these concerns. And soon, we will be launching the National Cyber Action Plan.ย But if AI gives those who could attack Britain new capabilities, significant new capabilities, we must use the same technology to make Britain harder to attack.”
He announced that he and the DCMS (Department for Digital, Culture, Media and Sport) Cyber Minister, Baroness Lloyd of Effra; and Dan Jarvis will lead on ‘the urgent use of frontier AI capabilities to find weaknesses within Government systems before attackers exploit them, protect essential government systems and respond more effectively to emerging threats. I want Britain to become one of the hardest countries in the world to attack using artificial intelligence.’ He said the UK Government will ‘now develop robust ways for government to respond to agentic AI incidents: looking in particular at how incidents can be contained, and how those affected can be supported.’
“I want cybersecurity discussed in every boardroom in Britain, rather than simply delegated to middle management,” he said. For the speech in full visit gov.uk.
Comments
Josh Langley, CIO at the data storage company Iron Mountain, said: ” This Cybersecurity Awareness Monthโs theme, โDonโt Make It Easy for Themโ, is a useful call to action. It means thinking beyond the moment an attacker gets through the front door. The AI Minister’s warning is just another high-profile reminder of why damage limitation is paramount in 2026.
“An assume-breach mindset asks a harder question: if a cybercriminal makes it into your environment, how much data can they steal, encrypt or destroy? Data resilience has a critical role to play here. Backups and recovery systems cannot be treated as an afterthought. If they remain connected to the same environment as production data, they can become another target for ransomware and data destruction.
“Critical data needs layers of protection, including isolation and immutable copies that attackers cannot alter or encrypt. Just as importantly, organisations need to regularly test whether those copies can actually be used to recover. Keeping attackers out will always matter. But true cyber resilience means planning for what happens when they get in, and making sure your most important data is still there when you need it.”
Tim Mackey, head of software supply chain risk strategy at the platform Black Duck, said: โOrganisations from the smallest enterprise to the largest governmental entity have built cybersecurity programmes around the premise that the attacker was a set of humans defining an attack around human values โ cash, control, and influence. This model allowed defenders the luxury of ignoring some cybersecurity gaps while focusing on others. AI agents donโt share human values โ they are purely goal oriented. While itโs easy to define guardrails for AI agents, those guardrails are both reactionary and clearly indicate to an AI agent that something useful in their goal-oriented quest might be behind the guardrail.
“Risk reviews by LLM providers are a key component of emerging AI risk strategies but they are no substitute for proper risk assessments of enterprise assets coupled with a focused effort to address identified weaknesses. After all, the one conclusion we can draw from the latest AI agent breaches is that AI agents are tenacious in solving their assigned goals and they donโt have an appreciation for collateral damage.โ
