A compromised agent can complete hundreds of actions before a person notices suspicious behavior. A kill switch provides a way to halt its operations and cut off access to connected systems. Build one before deploying an agent to production, rather than relying on investigation alone to stop it. section 4.8.8
A basic switch revokes the credentials an agent uses to reach internal systems. For a self-managed agent, it can also shut down the running agent. Where actions pass through a control plane, the agent can be stopped centrally instead of restricting access to each system separately. Token invalidation or network blocking should be controlled from a point the agent cannot reach. section 3.8 section 6.7.1
Combine the switch with controls that detect or limit trouble:
- Use automated triggers, such as an anomaly threshold that revokes credentials, alongside a manual override for the security team. section 4.8.8
- Set limits on spending, repeated attempts, and loops so an agent can be stopped without waiting for human detection. section 3.8
- Monitor actions in real time and identify agents that bypass the control plane. Logs help reconstruct incidents, but recording an action alone does not prevent it. section 3.8 section 4.8.8
Stopping a suspicious agent
An anomaly detector crosses a predefined threshold and automatically revokes the agent’s credentials. The agent loses access to connected systems while the security team investigates its recorded decisions; the team can also activate a manual override. section 4.8.8 section 3.8

