AI kill switch glossary and FAQ

Key facts

Sandbagging is strategic underperformance on an evaluation; detection methods tested by the UK AI Security Institute produced mixed results. Source
Independent verification organizations were introduced as attestors of shutdown capability by California's executive order of 18 September 2026. Source
No general statutory requirement for an AI kill switch was in force as of September 2026. Source

Glossary

Terms to know

Containment: Isolation of a system from networks, tools, data, or compute after a safety event.
Fail-safe: A design that defaults to a safer state when controls, communications, or authorization fail.
Frontier model: A highly capable general-purpose AI model near the leading edge of current performance.
Human override: A mechanism allowing an authorized person to interrupt automated operation.
Sandbagging: Strategic underperformance by an AI system on an evaluation, so that measured capability falls below actual capability. Relevant to kill switch design because nearly every proposed trigger is keyed to evaluation results. See: Evaluation integrity
Independent verification organization: An outside body that tests and attests to a developer's shutdown capability instead of relying on self-certification. Central to California's September 2026 executive order.
Sandbox: A restricted environment intended to limit what software or an AI system can access.
Tripwire: A measurable condition that triggers review, restriction, or shutdown.

FAQ

Common questions

What is an AI kill switch?

An emergency mechanism intended to pause, isolate, or shut down an AI system when it crosses defined safety thresholds.

Is it a single physical button?

Not usually. Proposals combine software controls, access revocation, compute shutdown, network isolation, and human authorization.

Could advanced AI evade it?

That is a central concern. Copying, deception, distributed deployment, and unclear trigger conditions may weaken shutdown controls.

Who should control it?

Proposals divide authority among developers, independent auditors, infrastructure providers, and governments. No consensus exists.

Who verifies that a kill switch works?

California's September 2026 executive order points to independent verification organizations: outside bodies that test and attest to a developer's shutdown capability rather than relying on self-certification. The model is new and untested at scale.

Study guide