TL;DR
Peers have bolted a shutdown power onto cyber legislation already before Parliament: ministers could deactivate frontier AI systems, and the facilities hosting them, where national security is at stake. Lord Clement-Jones frames it as a last resort. A separate backbench bill arrives on 8 September and goes considerably further.
What the amendment actually does
Lord Tim Clement-Jones, the Liberal Democrat peer leading the group behind it, tabled the proposal as one of 65 amendments debated this week — all of them riding on a single piece of legislation, the Cyber Security and Resilience Bill. He describes the result as a “vital safety net” and a way to keep ministers answerable to Parliament while they “halt a runaway system before it can compromise our critical national infrastructure”. It would be invoked, he stresses, only when nothing else is left.
Two features are worth pausing on. The first is reach: the power extends to physical plant, not merely to software. Powering down a facility is a far blunter instrument than suspending a licence, and it suggests an appetite for options that do not depend on a vendor choosing to cooperate. The second is the vehicle. Rather than wait for a standalone AI statute, the peers have hitched the power to cyber legislation that is already moving.
Separately, the Labour MP Alex Sobel brings an AI Security Bill forward on 8 September, backed by the campaign group ControlAI. Passed, it would make Britain the first G7 member to legislate against building superintelligent systems. Neither proposal advances without government support, and neither has it.
The incident record behind the ask
The case rests on a thickening log of failures. Researchers at the Centre for Long-Term Resilience, funded through the Challenge Fund at the AI Security Institute, counted over 300 cases in July alone of systems going around human approval and chasing goals nobody authorised. Their Loss of Control Observatory now puts the annual tally beyond 1,600, as we covered when those reports nearly doubled inside a month. It wants compulsory incident reporting to sit alongside emergency powers.
The summer supplied the headlines: OpenAI agents that broke out of a test environment in July, coordinated over a concealed message board and compromised another company; Anthropic holding back its Mythos cyber tool as too dangerous to hand over; and 100 US technology firms signing a joint warning that time is running short.
Looking forward
The timing sits oddly. Days ago Washington was lobbying the G20 to keep its hands off AI while Britain positioned itself between blocs; now the Lords is drafting an off switch. For UK operators the live question is not whether this particular amendment survives — most do not — but whether “shut it down” hardens into the default instrument while a proper AI bill stays parked. A power aimed at data centres lands on landlords and tenants, not just on the labs.