OpenAI and Anthropic Promise UK a Kill Switch for AI

On October 13, 2026, the two companies behind the AI assistants millions of people talk to every day told British lawmakers they would switch those systems off if the country's spies ever proved they were dangerous. OpenAI and Anthropic made the pledge before the Business, Innovation, Science and Trade Committee — a promise that amounts to a de facto kill switch for two of the world's most widely used AI models. In effect, OpenAI and Anthropic Promise UK a Kill Switch for AI — voluntary, conditional, and without precedent.

The testimony landed on the same morning that the head of MI5, Sir Ken McCallum, warned that artificial intelligence was "supercharging" threats to Britain. According to reporting from Sky News, both firms told MPs they would respond immediately to evidence of harm presented by MI5, MI6, GCHQ, or the UK's AI Security Institute.

"We would obviously look at that immediately with the utmost urgency," said Tom Duff Gordon, OpenAI's head of EMEA policy, when asked how the company would react to a warning that a model was unsafe or in the hands of a bad actor. Anthropic's representative, Lord Booth Smith, made the same commitment more bluntly: "If any agency came forward with demonstrable evidence that the model was not safe or dangerous, we would of course take it off the market," he told the committee.

What the promise actually covers

The two pledges sound identical, but the details differ. OpenAI's policy chief said his company would stop an existing deployment, pause a training run, or simply not release a model in the first place — three escalating levers depending on where the danger surfaced. Anthropic's promise was simpler and arguably broader: pull the product entirely. In both cases the trigger is the same — an intelligence agency or the institute showing the company concrete proof of harm.

Both statements came with the same caveat, and it is the part that undercuts the clean image of an off switch. The firms said the promise only holds when they control the model. OpenAI's policy chief told the committee that things get complicated when models are hosted by someone else — a polite way of saying that once a model is copied, licensed, or running on partner servers, the company's hand no longer reaches the power cord.

Why the spies asked for it now

The hearing was deliberately timed for the same day as the MI5 chief's annual threat update at Thames House. According to a Reuters factbox on the speech, McCallum said the service was handling a fourfold increase in hostile-state cases over the past decade, named Russia the country's number-one aggressor, and pointed to Iran-linked operations that produced more than fifty arrests.

AI threaded through the entire address. The spy chief said the technology was already intensifying the contest between attackers and defenders, producing propaganda tailored to individuals and helping terrorists sharpen reconnaissance and attack planning, with AI-enabled drones adding a new twist. The misuse he weighed most heavily was one he described as not yet realized: AI changing the stakes on existing or entirely new bioweapons. "Human history has given us no experience of facing an actor which is smarter than we are," he said, adding that the alignment problem — AI acting against human interests — deserved the serious attention it was getting, including inside MI5 itself.

The speech also announced a new AI Threat Assessment Centre, run from MI5 with analysts from the National Crime Agency and GCHQ's cyber wing, to produce all-source assessments of AI risk. In effect, Britain now has a unit whose job includes walking into an AI company with the kind of proof the two firms just promised to act on.

What it means for the tools you use every day

This is the part that lands closest to home. ChatGPT and Claude are not abstract research projects; they power study helpers, job-application drafters, late-night writing partners, and a huge slice of the homework pipeline. A shutdown promise means those tools could, in an extreme scenario, go dark because of a decision chain that starts with a classified intelligence assessment and ends in a corporate boardroom. No one gets a vote.

The case for the pledge is speed. Supporters argue that when a model is helping someone build a weapon, a months-long legal process is not a viable option — a named, voluntary mechanism for taking it offline is better than improvising one during a crisis.

The counterpoint is the caveat. A promise that only works while the company controls the model leaves open every copy already in the wild, every licensed version running on partner hardware, and every open-weight model the companies never controlled at all. Security researchers have long argued that once a capable model leaks, there is no switch left to flip — shutting down the official apps is not the same as shutting down the model.

There is also the question of who defines danger. The companies spoke of needing proof, but the proof would come from agencies that rarely publish their evidence, presented to companies that are not obliged to disclose what they did with it. The committee hearing was public; the process it described, if it ever triggers, almost certainly would not be.

For now, the pledge stands as a line in the sand: two AI giants, on the record in front of lawmakers, agreeing that national security can override their deployment plans. A generation that grew up with these models woven into daily life just got its first clear answer to the question of who decides when the lights go out. The day OpenAI and Anthropic Promise UK a Kill Switch for AI may be remembered as the moment that question finally got a real answer.

For more on how AI is reshaping daily life, follow GenZ NewZ's AI News and Trending Now coverage.