A runtime kill switch is a control that can terminate an AI agent session immediately, either manually or automatically when a policy is triggered. It acts like an emergency stop for agentic software that may be connected to files, APIs, cloud services, or internal tools.
In cyber security, this matters because AI agents can make decisions and perform actions faster than a human can react. If an agent behaves unexpectedly, exceeds its privileges, or begins touching sensitive systems, a kill switch can reduce dwell time and limit blast radius. In practice, defenders use it with identity controls, logging, and approval workflows so they can discover an agent, monitor its behavior, and shut it down if it violates policy. Attackers may also try to avoid or disable such controls, which is why the switch must be enforced through a trusted control plane rather than inside the agent alone.