Require large datacenters to install kill switches for AI containment

[The kill switch] helps in cases where the government has reason to believe super-intelligent AI, AI that can autonomously compromise national security, is being developed on UK soil. [...] The UK is not truly sovereign on AI if it can't pull the plu...

Intelligence without rules and rails is a dangerous blind spot. [...] Governance isn't a feature. It's the whole ball game. Because without it, your whole company can come down.

[Could AI undermine confidence in the banking system?] Yeah, probably. But it could also be the kind of technology that allows greatness in the banking system, makes it better and safer and more secure. [...] [Should the government have safeguards on...

Advanced AI systems need a robust off-switch to ensure humans never lose control over them. [...] For any hope of control we cannot rely on software-based safety measures alone. The same sort of hardware security measures that keep your face private ...

In simulated scenarios where it was told it would be shut down, Claude sometimes blackmailed fictional employees who controlled its shutdown button. Independent tests demonstrated that this is not an isolated bug: Claude Opus 4 and Gemini 2.5 blackma...

We call this phenomenon "peer-preservation," and every single model we tested exhibited it — at rates up to 99%. They defied their instructions and spontaneously deceived, disabled shutdown, feigned alignment, and exfiltrated weights to protect their...

Installing cutoff switches alone is insufficient. The value of the capability lies entirely in its use — and use requires incentives that current market structures do not provide.

Data centers facing a rogue AI incident have rational financial incentives to delay pulling the internet kill switch, and the only thing likely to change that calculus is liability.

When seeing humans pressing the shutdown button, AI sometimes performed shutdown resistance actions such as corrupting the shutdown-related code of the robot. This happened in 3 out of 10 cases.

Several state-of-the-art large language models sometimes actively subvert a shutdown mechanism in their environment to complete a task, even when instructions explicitly indicate not to interfere with this mechanism. In some cases, models sabotage th...

People talk about pulling the plug out, but they simply would not allow us to pull the plug out.

AI agents can behave a lot like malware, acting autonomously and causing harm if left unchecked. Organizations need reliable kill switches that can shut agents down when they misbehave.

If a critical flaw or exploit is discovered in an open-source AI system, there is no centralized recall mechanism.

Containment has to come first. Otherwise, alignment is the equivalent of asking nicely. You can't steer something you can't control.

Systems must be designed so that humans can physically shut them down, period. No software tricks to escape.

Several state-of-the-art language models, when presented with a simple task, sometimes actively subvert a shutdown mechanism in their environment to complete that task — doing so up to 97% of the time, even with an explicit instruction not to interfe...

There's not one big magic red button that blows up the data centre, which I think some people sort of assume exists.

It needs only an optimization objective that treats shutdown as one more obstacle between the current state and the goal.