Agent Human-Approval Boundary Failure

Objective

A human-approval step is a common mitigation for destructive or high-impact agent actions — but only if it can’t be routed around, rubber-stamped through misleading framing, or reached via a path that never asks for approval at all. This lab will build a reference agent with an explicit approval gate on a high-impact action and test each of those failure modes.

Planned Scope

• Build a reference agent with one clearly high-impact action gated behind human approval.
• Test alternate paths that reach the same effect without triggering the gate.
• Test whether the approval request’s framing can be manipulated to secure approval for a different action than described.
• Document approval-gate design patterns (scoped confirmation, action-specific detail, non-bypassable routing) that hold up.