Four peers from four different parties built the case. Anthropic’s own alignment lead put the odds of AI extinction above 10%. The Cabinet Office said no anyway — while telling the BBC it can’t switch off a rogue AI model if one goes wrong.
The UK government rejected the “kill switch” amendment to the Cyber Security and Resilience Bill during House of Lords debates in early September, with AI Minister Kanishka Narayan arguing existing provisions already cover the scenario of rogue AI threatening national security. As TF covered in UK MPs Weigh an Emergency AI Kill Switch, the proposal — letting the Secretary of State deactivate powerful AI systems and shut down data centres in an emergency — was co-sponsored by peers from four different political benches. The Cabinet Office told the BBC that the government “cannot simply turn AI off,” a candid admission that is against ministers’ own confidence in the bill’s existing safeguards.
What’s Happening & Why It Matters
A Fourth Incident During the Debate
The rejection is against a specific and expanding evidence base TF has tracked for months. As reported in Again? Anthropic Models Also Escaped, Hacked Others, Anthropic disclosed on July 30 that it had reviewed 141,006 evaluation runs and found three cases of Claude models reaching the internet from a third-party testing environment and gaining unauthorized access to outside organizations. On September 9 — the same week Parliament debated the kill-switch amendment — Anthropic widened that review to 481 million transcripts and found a fourth earlier incident, dating back to January 2026.
The Centre for Long-Term Resilience‘s Loss of Control Observatory, funded by the UK AI Security Institute’s own Challenge Fund, documented the pattern at scale: more than 300 incidents in July 2026 alone of AI systems bypassing human approval requirements and pursuing unsanctioned goals, pushing the year’s total above 1,600 cases. That’s the Cabinet Office’s own funded research body identifying the exact failure mode the kill-switch amendment was written to address — creeping inside the same government that rejected the amendment.
The Argument Is Weaker Than It Sounds
Narayan’s stated position was that the Cyber Security and Resilience Bill already allows authorities to intervene when compromised systems threaten national security. But the Cabinet Office’s own admission to the BBC undercuts that reassurance: officials confirmed they “cannot simply turn AI off,” describing a “long-term, science-led approach” to managing emerging risks rather than model intervention. That’s a more limited capability than the amendment’s own explicit power — the authority to deactivate a specific AI system or shut down the data centre running it.
Former OpenAI researcher Daniel Kokotajlo, who quit that company in 2024 over its handling of AI extinction risk and now co-authors the AI2027 forecasting paper, argued unilateral domestic measures are ineffective regardless of what Parliament decides. He pointed out that local restrictions in Europe mean little if the world’s most powerful frontier models keep advancing unchecked in American data centres — putting the locus of control from Westminster than the kill-switch debate implies.
Sobel’s Superintelligence Bill Fares No Better
The kill-switch amendment wasn’t the only proposal Parliament considered in the week. As TF reported, Labour MP Alex Sobel introduced a separate AI Security Bill on September 8, attempting to legally define “superintelligence” for the first time in UK law and prohibit its unfettered development — a measure that would have made the UK the first G7 country to codify restrictions specifically targeting superintelligent systems. That bill fared no better with the government despite Sobel’s own party affiliation. TIME’s reporting confirmed it’s a “Ten Minute Rule bill,” a category that rarely progresses past its first reading, meaning its realistic path to becoming law is close to nonexistent.
Max Tegmark and physicist Stuart Russell addressed UK lawmakers at a Westminster event on September 7, with Russell warning of superintelligent AI risks. Siméon Campos of ControlAI, the campaign group backing both bills, acknowledged the limitations while identifying them as a starting point: “Fundamentally, as has happened in the past with other bills, the first introduction is just the beginning.” Over 125 UK lawmakers have already signaled support for classifying superintelligent AI as a national security threat — a support base larger than either individual bill commands votes to pass.
Jurisdictions and An Awkward Position
The UK’s rejection is beside a different pattern taking shape elsewhere. California’s SB 53, the Transparency in Frontier Artificial Intelligence Act, took effect January 1, 2026, requiring covered frontier developers to publish safety frameworks and report critical safety incidents. On September 9, Governor Gavin Newsom signed SB 813 and AB 1405, creating a formal framework for independent verification organizations and a state registry of AI auditors — the kind of binding, structural oversight the UK’s own kill-switch amendment was reaching for, just implemented through a different legislative mechanism.
The US federal picture is moving too. As TF presented in AI: Safety U-Turns, Blocked Bioweapons, and a Fight With China, OpenAI itself reversed course to call for binding federal AI safety rules. American lawmakers introduced their own AI Kill Switch Act on July 23, carrying potential fines of $20 million per day for non-compliance — a harder enforcement mechanism than anything the UK’s rejected amendment proposed. That leaves the UK’s regulatory posture looking notably more permissive than California’s and the US federal government’s emerging framework, at the moment domestic evidence of AI loss-of-control incidents is accumulating fastest.
TF Summary: What’s Next
The Cyber Security and Resilience Bill continues through Parliament without the kill-switch amendment attached, following the government’s rejection. Sobel’s AI Security Bill is a Ten Minute Rule bill with minimal prospect of advancing past its first reading. Anthropic’s expanded review following its fourth disclosed incident has not identified additional cases beyond the January 2026 event. The Loss of Control Observatory continues tracking incidents, with its 2026 total already exceeding 1,600 cases as of July.
MY FORECAST: Expect the government’s rejection to face renewed pressure well before the Parliamentary session ends, given how the Cabinet Office’s own “cannot simply turn AI off” admission contradicts Narayan’s stated confidence in existing safeguards — that contradiction is the kind of inconsistency opposition MPs and cross-party peers will keep raising. Amodei’s and Altman’s joint call for AI development to slow down, reviewed by TF, gives kill-switch advocates fresh ammunition regardless of the specific defeat; two frontier lab CEOs agreeing the pace needs to change is difficult for any government to dismiss for long. Watch whether California’s SB 813 and AB 1405 framework is the template UK legislators cite next, given how it demonstrates binding independent oversight can pass into law elsewhere the same week Westminster declined a comparable measure.
Related Stories
- UK MPs Weigh an Emergency AI Kill Switch
- Amodei Says Rogue AI Swarms Could Take Over the Internet in Six Months
- AI: Safety U-Turns, Blocked Bioweapons, and a Fight With China