UK lawmakers push for AI ‘kill switch’ powers
Written by Rebecca Uffindell Fri 4 Sep 2026
The UK Government could be given powers to shut down data centres or deactivate AI systems as a last resort when they pose a threat to national security, under proposals being considered as part of the Cyber Security and Resilience Bill.
The amendment was proposed by Liberal Democrat peer Lord Tim Clement-Jones as the Bill progresses through Parliament. The BBC, which first reported the proposal, quoted Clement-Jones describing the measure as a “vital safety net” that would provide a democratically accountable means to halt a runaway system before it could compromise critical national infrastructure.
According to the BBC, the powers would be intended for use only as a last resort.
Their potential application to data centres is significant. The UK designated data centres as Critical National Infrastructure (CNI) in 2024, placing the sector alongside infrastructure including energy and water in recognition of the essential services it supports.
New Cyber Bill Could Introduce Emergency Powers
The proposal would sit within the wider Cyber Security and Resilience Bill, which is intended to strengthen protections around the digital services and infrastructure on which the UK depends.
Clement-Jones’ amendment raises a more specific question about the wave of progressively capable AI: what authority should the government have if an AI system itself becomes a national security threat?
AI ultimately depends on physical infrastructure. Advanced systems require data centres, compute, networking, and electricity, meaning an emergency intervention aimed at an AI system could also require action against the infrastructure supporting it.
Simon Edwards, CEO of SE Labs, said increasing AI autonomy made those questions more pressing.
“As AI is becoming increasingly autonomous, it is posing a huge threat to individuals, and businesses of all sizes, on top of our critical national infrastructure,” he said.
Edwards argued that a dedicated national cyber strategy and adequate testing of software and services for vulnerabilities should also be priorities.
The proposed kill switch is not, however, the only attempt to bring greater AI security controls into UK legislation.
Alex Sobel Plans Separate AI Security Bill
Labour MP Alex Sobel has said he plans to introduce a separate AI Security Bill in Parliament on 8 September, supported by campaign group ControlAI.
ControlAI says its work focuses on informing policymakers and the public about risks associated with superintelligent AI. According to the organisation, it has briefed 375 lawmakers across the US, UK, Canada and Germany, including the UK Prime Minister’s office, since November 2024.
The initiative is separate from Clement-Jones’ amendment to the Cyber Security and Resilience Bill, but it reflects a related concern: whether existing legislation provides sufficient safeguards as AI systems become more capable and autonomous.
This separation is highly relevant. The emergency powers proposed through the Cyber Security and Resilience Bill concern what government could do when an AI system or associated infrastructure presents an extreme threat. The AI Security Bill raises the question of what controls should exist around the development of highly capable systems before that point is reached.
Enterprise AI Faces a More Immediate Control Problem
The principle of placing limits on autonomous systems also has implications inside businesses.
“The Bill is right that some AI capabilities need hard limits. But the same principle must apply inside businesses, not only to the behaviour of frontier models,” said Andrea Sivieri, Chief Product and Technology Officer at CoreView.
Sivieri stressed that many organisations currently lack sufficient visibility over what their AI agents are permitted to do, who authorised that access, or how permissions have changed.
An agent with persistent access to an enterprise environment can effectively become a privileged user. In Microsoft 365, for example, sufficiently privileged access could allow an agent to change permissions, create, or delete users, or alter security settings.
“Agents should only have access to the data and systems they genuinely need,” said Sivieri.
He argued that organisations should maintain records of what agents access, change, send, or delete, alongside clear rules determining who approves their permissions and how regularly those permissions are reviewed.
The enterprise problem is considerably narrower than the extreme national security scenarios contemplated by lawmakers. But both raise a common issue: as AI systems gain greater autonomy, organisations need to establish where their authority ends and retain mechanisms capable of enforcing those boundaries.
Hugging Face Attack Showed Increasing AI Autonomy
A cyber incident disclosed by Hugging Face in July provides an example of how autonomous AI capabilities are already changing the threat landscape.
The company said an intrusion into part of its production infrastructure had been conducted “end to end” by an autonomous AI agent system, describing the incident as different from those it had previously handled.
According to Hugging Face, the attack began through its data-processing pipeline before the actor escalated access, harvested cloud and cluster credentials and moved laterally into internal clusters.
“The campaign was run by an autonomous agent framework […] executing many thousands of individual actions across a swarm of short-lived sandboxes,” said the company.
Hugging Face’s subsequent forensic analysis covered more than 17,000 recorded events from the attacker’s action log. The company said the large language model used by the attacker remained unknown.
The incident should not be characterised as an example of superintelligent AI, nor as a runaway model independently deciding to attack Hugging Face. It was an offensive cyber campaign conducted using autonomous AI tooling.
Its relevance to the wider debate lies in what it demonstrates about increasing autonomy. AI systems can already execute complex sequences of cyber activity rapidly and at significant scale, providing a tangible indication of why policymakers and security specialists are considering how more capable autonomous systems could be controlled.
Cyber Resilience Will Have to Operate at AI Speed
The growing speed of AI-enabled activity creates a corresponding challenge for defenders.
“Attacks now move at AI speed, disruption moves at AI speed, and complexity grows at AI speed, so if your resilience doesn’t move at AI speeds, you’ve already lost,” said Andy Ward, SVP International at Absolute Security.
Ward argued that organisations needed stronger AI-powered cyber resilience strategies and real-time visibility as the speed and complexity of attacks increased.
“Without robust AI-powered cyber resilience strategies and real-time visibility in place, the UK risks sleepwalking into deeper vulnerabilities,” added Ward.
That places emergency shutdown powers at one end of a much broader resilience spectrum.
A government kill switch would be intended for an exceptional situation. Before that threshold is reached, organisations still need the ability to detect abnormal AI behaviour, understand which systems and data are affected, contain incidents, and recover operations.
The rising autonomy demonstrated by incidents such as Hugging Face makes the time available to make those decisions potentially much shorter.
The AI ‘kill Switch’ Debate is Becoming International
The concept is not emerging in the UK alone. In the US, lawmakers have also introduced legislation explicitly framed around an AI kill switch.
Congressmen Ted Lieu and Nathaniel Moran, a Democrat and Republican respectively, introduced the AI Kill Switch Act, adding a bipartisan dimension to the debate over mechanisms for constraining advanced AI systems.
The US proposal followed growing concern over AI systems’ ability to perform security-sensitive actions, including OpenAI’s disclosure of model behaviour involving a major repository of computer code.
The UK and US proposals are not identical and should not be treated as components of a coordinated legislative effort. But their emergence reflects a similar policy problem.
AI governance has largely concentrated on how systems should be developed, tested, and deployed. Increasing autonomy is now adding another question: what happens when those safeguards are insufficient?
For the UK, that question has particular significance because the infrastructure supporting advanced AI includes data centres that are themselves designated as critical national infrastructure.
The powers being considered would be exceptional and, according to the BBC’s reporting, reserved for a last-resort scenario.
But the emergence of kill-switch proposals on both sides of the Atlantic shows that policymakers are beginning to think beyond rules governing normal AI operation.
As systems become more autonomous, the debate is now more often extending to the point at which intervention becomes necessary — and who should have the authority to stop an AI system when the risks become unacceptable.
Written by Rebecca Uffindell Fri 4 Sep 2026
Tags:
AI Governance Cyber Security and Resilience Bill UK GovernmentMost Viewed News
Fri 2 Oct 2026
AI skills gap grows despite 64% of workers using AI