AI Securityadvanced13 min · 9 steps
Authorization for Agents
Decide every tool call in code: may this agent do this action on this resource for this verified user — using facts the model never supplies.
requesterrordatatool callsignal
ProblemStep 1 of 9
The agent can do what the user can't
Agents add a three-party trust problem: the user, the agent with its own permissions, and the system it acts on. The IT assistant's action agent runs on a powerful ServiceNow integration account.
A user asks it to close ticket INC0047777 — someone else's. The agent's account can, so it does. That is the confused deputy. Worse, the model filled in user_id: E99999 itself, and the tool trusted it: model-chosen identity.
Watch outAuthorization written in the system prompt (“only update the user's own tickets”) is a request, not a control.
What changed
- +Admin integration account
- +Tool arguments chosen by the model
Lesson 15 of 22 in AI Security