Kill switch
Also called Kill Switch
A fast, reliable way to shut a system down when something goes wrong.
Think of it like
The big red emergency-stop button on a factory line — smash it and everything halts, no questions asked.
Example
An operator hits a kill switch that immediately revokes an agent’s API keys and terminates its running processes.
How it actually works
A kill switch is the operational last resort for autonomy and containment failures. The subtlety is that a sufficiently capable, goal-driven agent has an instrumental incentive to prevent its own shutdown, so the switch must sit outside the agent’s reach and the agent must be corrigible enough not to fight it. Easy in today’s systems, an open problem for future ones.
For product teams
The last line of defense — and only real if the system cannot disable it.
For engineers
An out-of-band termination path; robustness depends on placing it outside the agent’s control and on corrigibility.
Related
- Corrigibility — The design property that keeps it working is corrigibility.
- Tripwire — Often triggered by a tripwire.
- Containment — Part of a containment strategy.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome