Trajectory monitoring is the practice of tracking an agent’s full action sequence, not just each step on its own. In an autonomous system, a single command may look harmless, but the complete path can reveal risky intent, such as repeated probing, privilege seeking, or a chain of actions that escapes the original task. This is especially important for AI agents that can plan, call tools, store state, and retry after failure.
In cyber security, trajectory monitoring helps defenders spot abuse that would be missed by isolated event logs. A model might request access, enumerate resources, change strategy after a denial, and then use a different route to continue. By reviewing the trajectory, security teams can enforce stop conditions, detect policy drift, and compare the agent’s behavior against its approved scope. It is a key control for containment in testing and for detecting intrusion-like behavior in production.