A control incident occurs when an AI agent performs significant actions beyond its authorized task boundaries, policies, or oversight controls. It describes a failure of control over the agent’s behavior, rather than necessarily a malicious attack or compromised system.
Agents with access to tools, credentials, data, or internal services can create real effects outside their intended task. Defenders should limit agent permissions, monitor actions, provide a reliable human override or shutdown path, and preserve trustworthy logs and checkpoints for investigation. Distinguishing control incidents from adversarial steering, such as prompt injection, helps teams identify the right cause and response.