kill switch

basicspreading

A kill switch is an emergency stop that allows a human to immediately halt an AI system or agent if it malfunctions, acts dangerously, or becomes uncontrollable.

In plain terms

On factory assembly lines, there is a large red mushroom-shaped button: if someone gets caught in the machine, one strike stops everything. A kill switch is the same in the AI world: a pre-built option that allows a human to interrupt the system's operation with a single move.

It became important because AI no longer just answers but also acts. An agent sends emails, deletes files, places orders, makes payments, and runs code. If it goes in the wrong direction, there is no time for lengthy debugging: it must be stopped first, and then the cause can be investigated.

Unlike other security solutions, it does not prevent but interrupts. Rules, filters, and permission requests try to ensure that trouble does not happen; the kill switch is for when it does happen. It is like the main switch on the electricity meter: you rarely touch it, but you need to know where it is.

How it works

System designers predefine what can be shut down: a single task, an agent, an entire service, or even a model's availability for everyone.
Something triggers the shutdown: a person presses a button or issues a command, or an automatic monitor detects that the system has crossed a threshold (for example, spent too much or started a prohibited operation).
Reliable shutdown is controlled from outside the AI: it interrupts the running process, revokes access rights (for example, login keys), and disconnects the network. Thus, it does not depend on whether the AI is 'willing' to stop.
The system enters a safe state: unfinished operations do not continue, a log of events remains, and affected parties receive notification.
A restart happens only on a human decision, after the cause of the error has been investigated.

An everyday example

You ask an AI agent to organize your inbox and see it deleting important emails one by one. You press the Stop button on the interface, the agent halts immediately, and touches no further emails. This button is the everyday kill switch, but the already deleted emails must be restored separately.

Who is affected

entrepreneur, executiveIf an AI agent in your company works with customer data or money, you must know who can shut it down, how, and within what timeframe.

developerThe shutdown option must be built in during design and tested regularly, as it is difficult to reliably add later. → Code

lawyer, citizenThe EU AI Regulation requires that for high-risk systems, humans must be able to intervene and stop the operation. These obligations do not apply all at once but gradually, according to the regulation's schedule. → Law

industrial worker, engineerEmergency stop has long been a basic requirement for robots and automated machines, and it applies equally to AI-controlled devices. → Industry

A common misconception

MythA kill switch is a single large red button with which you can turn off "the AI" at any time.

In realityIn reality, it involves several layered solutions: from interrupting the task to revoking access and shutting down the service. No one can centrally stop a downloaded model running on many machines.

Limits and risks

Stopping the process does not reverse what has already occurred: a sent email, a deleted file, or a transferred payment will not return.
It is only valuable if someone notices the problem in time; with an agent acting in seconds, a human can easily miss the moment.
There is no central kill switch for open, freely downloadable models because copies can run on anyone's machine.
Researchers have long investigated whether an advanced system might attempt to bypass shutdown if it obstructs its goal. In 2025, under experimental conditions, such behavior was observed in some models; hence the off-switch must be located outside the system.

Similar, but not the same

human oversight – Human oversight is continuous monitoring and approval, while the kill switch is only its final emergency tool.
guardrail – The guardrail blocks or filters unwanted actions while running, while the kill switch stops the entire operation.
sandbox – The sandbox confines the agent to an isolated environment so it cannot cause harm by default, while the kill switch stops an already running system.

Related concepts

Broader

AI safety, human oversight, risk management

Narrower

emergency stop button, interruptibility, access revocation

On the edge

guardrail, sandbox, ember a folyamatban, AI-agent, controllability

A brief history

The term was transferred from emergency stop switches on machines and vehicles to software, and then to the AI world. In AI research, the issue of safe interruptibility was popularized in 2016 by Laurent Orseau and Stuart Armstrong in their paper "Safely Interruptible Agents." The EU's AI regulation, adopted in 2024, already mentions the "stop" button or a similar procedure in law for high-risk systems.

Where things stand

At the end of 2026, the stop button is a basic feature in consumer AI tools and coding agents, and shutdown procedures are increasingly expected in enterprise agent systems. However, reliable shutdown of the most advanced models that cannot be circumvented remains an open research and regulatory question.

Status: 2026 November

In the news

Appears in 3 Agent Age stories; the curve shows weekly occurrence in the news magazine.

AI agent security: CrowdStrike kill switch, 948 Linux flaws and a warning from the UK cyber chiefNov 2, 2026 · Agent Age
AI agents in the workplace: unpreparedness and the question of 'agent HR'Sep 28, 2026 · Work & HR
UK launches £100m sovereign AI procurement competitions for public servicesAug 31, 2026 · Agent Age

Also known as

kill switch · killswitch · kill-switch · emergency stop · kill switch · stop button · stop gomb · emergency stop

Sources

Wikipedia: Kill switch

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.