What is AI kill switch?

What it is

An AI kill switch is an emergency stop for software that acts on its own. When an AI agent is reading files, calling tools and making changes without a person approving each step, someone needs a reliable way to halt it fast. The kill switch is that way.

What a good one does

A kill switch is more than closing a window. A useful design covers several things:

• Stops the running work. The agent's current task ends, not just its chat window.
• Prevents a restart. The agent cannot quietly relaunch itself or be relaunched by another process.
• Cuts access. Network connections, credentials and tool permissions are revoked so a stopped agent cannot keep acting through side doors.
• Leaves a record. Who pressed the switch, when, and what the agent had been doing are logged for later review.

Where it sits

A kill switch works best as the last layer in a stack. Earlier layers keep problems small: an agent sandbox limits what the agent can touch, AI guardrails filter inputs and actions, and AI observability shows what the agent is doing. The kill switch handles the case where those layers were not enough.

Why it matters

Agents can run for a long time, spend money and act across many systems. A mistake or an attack, such as prompt injection, can turn a small error into a large one before anyone notices. The practical question for any team running agents is how many seconds it takes to stop one, and who is allowed to do it.

An example

Picture an agent told to clean up old files that starts deleting the wrong folder. An administrator uses the kill switch: the process ends, it cannot relaunch, its access to the file server is removed, and the log shows exactly which commands ran. The damage is limited to what happened before the stop.

What to check

Before you deploy an agent, decide who holds the switch, test it, and confirm it works without needing the agent's cooperation. A stop button that depends on the agent agreeing to stop is not a kill switch.