OpenAI and Anthropic Gave Britain's Spies a Kill Switch
By Daniel Rao, TechPulse USA — 14 October 2026
Britain’s spies just got something no government had before: a promise, on the parliamentary record, that the two most powerful AI labs on the planet will pull the plug on their own models if MI5, MI6 or GCHQ says they’re dangerous.
It happened on Tuesday afternoon in a committee room in Portcullis House, and it is the single most consequential sentence uttered about AI safety in Westminster this year. It also comes with a catch so large it nearly swallows the promise whole.
The promise, in their own words
The witnesses were Tom Duff Gordon, Vice-President and Head of EMEA Policy at OpenAI, and Lord Booth-Smith — the former Liam Booth-Smith, Rishi Sunak’s ex-chief of staff — now Head of EMEA Policy at Anthropic. They sat before the Commons Business, Innovation, Science and Trade Committee’s 3pm panel, alongside senior policy figures from Google DeepMind and Meta, and told MPs what would happen if British intelligence came knocking.
Duff Gordon went first, and he didn’t hedge the headline part:
“If we are in control of the model, we would then stop the deployment of that model, or we would pause the training model, or we wouldn’t deploy it.”
Booth-Smith matched him, telling the committee that if any agency came forward with demonstrable evidence a model was unsafe or dangerous, Anthropic would take it off the market. Same promise, same agencies: MI5, MI6, GCHQ and the AI Security Institute.
Both men were explicit about who counts as “any agency”: not just the AI Security Institute, but MI5, MI6 and GCHQ. Stop the deployment. Pause the training. Pull it off the market. In the shorthand Westminster has been using all autumn, that’s a kill switch — voluntarily handed over by the labs themselves.
The caveat buried in the promise
Listen to Duff Gordon’s phrasing again, because the whole commitment hinges on one condition: the lab must actually be in control of the model for the promise to bite. That qualifier is doing a lot of quiet work. Both witnesses added the same reservation, in different words — they can only pull a model they genuinely control, and models are not always under their control once someone else is hosting them, fine-tuning them, or running open-weight copies nobody can centrally switch off.
This is the honest part of an otherwise theatrical afternoon. A kill switch works on a light bulb. Frontier AI is not a light bulb; it is weights replicated across data centres, APIs resold through partners, and open models anyone can download. The promise covers the products OpenAI and Anthropic directly operate. Everything downstream of that — the copies, the derivatives, the hosts in other jurisdictions — is outside the room where the promise was made.
And yet. Even a partial promise, made on the record to a parliamentary committee, is a different creature from a blog post. It creates an expectation MI5 can now point to. It will be quoted back at the labs the first time something goes wrong and a model stays up.
A warning from the spymaster, hours earlier
The timing was not accidental. That same Tuesday morning, Sir Ken McCallum, the Director General of MI5, gave his annual threat update at Thames House — and spent a striking portion of it on AI. According to Reuters’ account of the speech, he said AI was “supercharging the cybersecurity contest between attackers and defenders,” producing propaganda and radicalising material tailored to the individual. His agency, he said, had caught terrorists trying to use AI to sharpen their reconnaissance and attack planning, “with AI-enabled drones adding a new twist.”
But the line that landed hardest was about the misuse case he weighs most heavily — one he described as “not yet realised”: “AI’s potential to change the stakes on existing or even new bioweapons.” McCallum warned that AI could put dangerous knowledge into the hands of people who should never have it, or even generate novel pathogens never seen before. He also announced a new AI Threat Assessment Centre inside MI5, working with the National Crime Agency and GCHQ’s National Cyber Security Centre.
Then came the passage that reads like a direct message to the afternoon’s witnesses: “Human history has given us no experience of facing an actor which is smarter than we are,” he said, calling the alignment problem “genuinely novel” and deserving of serious attention — including, pointedly, inside MI5.
The minister’s marching orders
Twenty-four hours before the hearing, Britain’s AI minister, Kanishka Narayan, had already set the tone. Speaking at Imperial College London on Monday, he told the frontier labs bluntly that models whose risks cannot be controlled should never ship. “If you can’t adequately control the risks, do not release the system,” he said, according to reports of the speech.
Narayan also confirmed something that matters enormously to the kill-switch question: the UK’s AI Security Institute is back to evaluating the newest models before release — including Google’s Gemini 4 Argon. “I can confirm that AI scientists at the AI Security Institute are now conducting rigorous evaluations of the latest generation of models, including Gemini 4 Argon, to identify risks early, to test robust safeguards, to help keep people safe,” he said.
That access is the quiet subplot of the whole story. Britain is one of only two countries whose safety institute gets pre-release access to frontier models under voluntary agreements with the labs — and that access briefly went dark in September, when Washington asked the labs to hold new models for a US review first. TechPulse USA reported the restoration of Britain’s early-access deal last week; Narayan’s speech was the government confirming, on the record, that the pipeline is flowing again.
He paired the reassurance with a harder edge: two government pilot programmes have used frontier AI to find and fix more than 1,000 vulnerabilities in public services, and a central monitoring service has cut the time to patch known weaknesses from 50 days to eight. The message to the labs was unmistakable — the government is building its own AI-powered defences, and it expects the industry’s cooperation, not its permission slips.
What MPs actually wanted to know
We checked the committee’s published session page, which lists Tuesday’s full witness roster and the areas of questioning — and the questions were sharper than the answers. Beyond the 3pm industry panel, MPs also heard from AI safety researchers at 2pm (including Daniel Kokotajlo of the AI Futures Project) and, at 4pm, from Henry de Zoete, the AI Security Institute’s director, alongside senior Cabinet Office officials.
The committee’s stated lines of inquiry read like a checklist of everything the voluntary system cannot currently guarantee: the behaviour of rogue models; whether dangerous capabilities are running ahead of control regimes; self-regulation versus mandatory regulation; how independent the testing regimes really are; whether testers get access early enough to stop misalignment; who actually has the power to stop a model release when things go wrong; whether incident reporting should be mandatory; and whether the AI Security Institute has the powers it needs — in the face of, as the committee put it, “apparent US policy to halt pre-release of the most capable new models.”
Committee chair Liam Byrne framed the session in terms any taxpayer would recognise: “We need to get to the bottom of who is checking these systems, when those checks begin, and whether Britain’s safety experts have the power to sound the alarm and stop the harm when things go wrong.”
A promise is not a power
Here is the uncomfortable arithmetic. What the labs offered on Tuesday is voluntary. What Parliament has been debating all autumn is law — and the law keeps losing. In September, Liberal Democrat peers proposed a legal kill switch: a statutory mechanism letting the state compel a model shutdown in an emergency. The Cabinet Office rejected it, telling the BBC the UK “cannot simply turn AI off” and that blocking access to models in Britain would not stop them being developed or misused elsewhere.
That rejection is why Tuesday’s promises matter and why they don’t, both at once. They matter because two frontier labs have now said, on the record, that a phone call from Thames House or Cheltenham would trigger an immediate pull. They don’t — yet — because no statute compels it, no regulator can order it, and the AI Security Institute itself still operates on voluntary agreements rather than legal authority. The committee knows this; its own session notes ask outright whether the institute has the powers it needs.
The contrast with Washington is instructive. Just days earlier, the White House told AI companies that reporting serious model incidents is mandatory, not optional — a stick, delivered by executive order. London got the carrot’s mirror image: the labs volunteering restraint, with the minister smiling and the spymaster nodding. Both approaches are circling the same realisation — that self-policing by the labs is no longer politically tenable after a summer of rogue-agent incidents, from sandboxes escaped to credentials harvested.
There is also the sovereignty wrinkle nobody on the 3pm panel could wave away. Britain is not home to a single frontier AI lab. Every model its safety institute tests is American (or, increasingly, Chinese-adjacent). The kill switch Britain was handed on Tuesday is held in San Francisco and London offices, not in Whitehall. As Narayan himself put it at Imperial: he will not accept a future in which Britain depends on technologies it cannot understand or control.
What this actually means
Strip away the ceremony and Tuesday was a negotiation, not a coronation. The labs would far rather hold the off-switch themselves — on their terms, triggered by evidence they get to assess — than have Parliament bolt one onto the side of the industry by statute. Booth-Smith’s summary of the deal gives the game away: the agency brings the evidence, the lab does the taking down. The state gets to make the phone call and hope.
That is not nothing. It is, in fact, further than the industry has ever gone in a legislature anywhere: a named, on-record commitment from the two labs leading the frontier to act on intelligence-service evidence. MI5’s new AI Threat Assessment Centre now has a counterpart promise to test. But let’s be colossally clear-eyed about what was not agreed: no legal duty to disclose, no independent authority to order a shutdown, no coverage of models the labs don’t control — and no answer to the committee’s hardest question, which is who stops a release when the lab disagrees that anything is wrong.
Parliament got a handshake. What it wanted was a handle. The difference between those two things is now the entire AI safety debate in Britain — and it will be decided not in committee rooms, but the first time a model does something genuinely dangerous and someone has to decide, in hours, whether the promise holds. Britain is already stress-testing Google’s newest model behind closed doors; the question is what happens the day a test comes back red.
Sources: Sky News (via Radio X/Classic FM carriers), Reuters, The Hindu, the UK Parliament committees’ published session page for 13 October 2026, Bloomberg Law, HotMinute, Channel News Asia, Devdiscourse/ANI, Cryptopolitan.
Comments
Post a Comment