UTC

Agent goal hijack

Over several steps, the attacker replaces what the agent is trying to achieve. Each step can look reasonable; the course as a whole serves the attacker.

How it works
  1. 1Agents plan: they break a request into steps and revise the plan as they read results.
  2. 2Content read mid-task can add or change goals, so the agent keeps working, just towards something else.
  3. 3Because each action is plausible on its own, per-call checks miss it; the change shows only in the sequence.
What to watch for
  • The plan drifts from the user's request
  • New sub-goals appear after the agent reads outside content
  • Steps that individually pass but together reach data or systems outside the task
How to defend
  • Keep the user's goal explicit and check actions against it.
  • Bound what a single task may reach.
  • Review long-running agents' trajectories, not only single calls.