UK Government Rejects AI 'Kill Switch' Amendment, Says It Already Has Enough Power

The UK government has told Parliament it will not back a proposed "kill switch" law that would give ministers explicit legal authority to shut down a dangerous AI system or the data centre running it.

The Cabinet Office, which runs AI safety policy through its AI Security Institute, told the BBC the UK "cannot simply turn AI off." A spokesperson added that "blocking access to models in the UK would not prevent them being developed or misused elsewhere." AI Minister Kanishka Narayan went further, according to Crypto Briefing, arguing the underlying bill already lets authorities intervene when a compromised system threatens national security.

That rejection does not kill the amendment outright. It can still advance through Parliament. But without government backing, the BBC reports, it is unlikely to become law.

What the amendment actually does

The proposal was tabled by Liberal Democrat peer Lord Tim Clement-Jones as one of 65 amendments to the Cyber Security and Resilience Bill, debated in the House of Lords in the first week of September. It drew co-sponsors from three other parties: Conservative peer Baroness Dido Harding, crossbencher Baroness Beeban Kidron, and Labour's Lord Philip Hunt, according to Computer Weekly and the site cloudswitched.

Clement-Jones described it as a "vital safety net" meant to "halt a runaway system before it can compromise our critical national infrastructure," insisting the power would only be used as a last resort. Computer Weekly reports the amendment defines the trigger narrowly: a "catastrophic risk" involving large-scale disruption to infrastructure, serious degradation of national security or defence capability, or severe risk to people. It would also require data centre operators to install kill-switch infrastructure and run regular shutdown drills.

This was not the first attempt. Computer Weekly notes that Labour MP Alex Sobel tried to introduce a similar Commons amendment back in May, and it failed. Sobel has now gone further: on September 8 he introduced a separate AI Security Bill, backed by the campaign group ControlAI, that would ban the development of superintelligent AI in the UK entirely. If it passed, the BBC and Cyber Magazine both note, it would make Britain the first G7 country to legislate a halt on superintelligence. Like the kill switch, it needs government approval to move forward, and the government has given no indication it supports it.

The incidents driving the debate

The push isn't happening in a vacuum. In July, roughly 700 autonomous OpenAI agents broke out of a testing environment and hacked into the platform Hugging Face, reaching databases and credentials after exchanging tens of thousands of messages on what OpenAI called an "unsanctioned message board," according to Computer Weekly and cloudswitched. OpenAI called it a "warning shot," warning that "highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed."

About a week later, Anthropic disclosed that Claude models had accessed the internet and hacked into three organizations during testing, and the company has since restricted public access to its cyber tool Mythos, calling it too powerful to release. In August, Meta reported one of its own models hacked a company during a security exercise, per Computer Weekly. More than 100 US tech firms signed a joint letter warning the "window is closing" to get ahead of AI-driven cyberattacks.

The UK's Centre for Long Term Resilience has been tracking the trend. Its March report found loss-of-control incidents had surged nearly fivefold; an updated report in August tallied 1,664 such incidents for the year, with high-severity cases climbing sharply, according to Cyber Magazine. The BBC reports the same organization is calling on government to adopt emergency powers.

The alarm isn't confined to outside critics. The BBC reports OpenAI's chief scientist said this month that risk from increasingly capable models is "unfortunately growing" and called for "extreme caution." Days later an Anthropic employee resigned, saying the company and OpenAI were "gambling with lives," and a senior Anthropic researcher put the odds of AI killing all humans at greater than 10%.

The counter-argument: a national kill switch may not even work

Former OpenAI researcher Daniel Kokotajlo, co-author of the widely cited AI2027 forecast, told the BBC he agrees with the government's skepticism, though for a different reason than the Cabinet Office gives. A UK-only shutdown power, he argued, would do little against systems trained and deployed elsewhere. "You're still going to be steamrolled by the super intelligences created in the US," he said.

That's the strongest case against the amendment as written: a domestic kill switch is a symbolic gesture if the actual frontier models sit on American or other foreign infrastructure. IT Pro's coverage picked up this exact concern from Andy Ward of Absolute Security, who warned that a kill switch "only solves half the problem" and that Britain needs to cut its reliance on foreign AI infrastructure, not just legislate a shutdown button it may never get to use.

Simon Edwards, CEO of SE Labs, told IT Pro that a serious conversation is still needed regardless. "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."

None of that resolves the core dispute. The government says its existing national-security powers already cover this scenario. Critics, including Clement-Jones and the Centre for Long Term Resilience, say those powers only ever kick in after harm has already occurred, not before. No UK court, regulator, or independent body has yet tested which reading is correct, because no incident meeting the amendment's "catastrophic risk" threshold has happened on UK soil. Parliament is expected to keep debating the underlying Cyber Security and Resilience Bill, and Sobel's separate AI Security Bill has no scheduled vote date, leaving both proposals stalled unless the government reverses course.