AI Securityadvanced13 min · 10 steps
Human Approval for Actions
Make high-impact agent actions wait for the right person to approve the exact change — bound to its parameters, single-use and short-lived — so approval is a control, not theatre.
requesterrortool calldata
ProblemStep 1 of 10
The approver approved what the model told them
A contractor asks the IT assistant for VPN access. A malicious ticket the agent read earlier has injected an instruction, and the approval message the model writes says “routine password reset”.
The manager clicks Approve. What actually runs is add_to_group(E12345, VPN-Contractors) — remote network access. An approval that can be shown misleadingly, bypassed, replayed or self-approved is security theatre.
Watch outIf the model writes the approval screen, the attacker writes the approval screen.
What changed
- +Approval message written by the model
- ~Approver sees a different action than the one that runs
Lesson 16 of 22 in AI Security