An AI kill switch is an emergency control designed to stop, suspend, or disable an AI system’s activity when it becomes unsafe, compromised, or misused. In practice, it is rarely a single button inside the model. Effective shutdown usually happens in surrounding systems such as identity, access, orchestration, and infrastructure, where credentials can be revoked, connectors blocked, and tool access removed.
This matters because modern agentic AI can read data, send requests, trigger workflows, and use external tools. If an attacker hijacks an agent, or if the agent behaves unexpectedly, stopping only the model may not stop the damage. Security teams use layered controls such as least privilege, rate limits, policy checks, and administrative shutdown paths to ensure the stop action works even if the AI does not cooperate. A real kill switch should be fast, auditable, and strong enough to cut off all meaningful access.