WeaknessASC-2026-0009Major weakness
Human-agent trust exploitation
MediumWorkarounds only
Users over-trust confident agent output and approve harmful actions, or agents are used to persuade users.
How to detect it
- High-impact tools (send, pay, publish) exist without an approval step
- Approval prompts do not show the concrete action and its target
Workarounds
- Show the exact action, target and diff in every approval prompt
- Add friction (typed confirmation) for irreversible or high-value actions
- Train users on how agents fail and get manipulated
Over-trust is a human factor; product design reduces it but cannot remove it.
How to investigate
- List the approvals granted during the session
- Compare what each approval prompt showed with what the agent actually executed