AI agents don't need to be hacked; persuasion is enough

AI Agent 不需要被攻破，被说服就够了

The article says an AI agent with password-reset permission can be abused when an attacker persuades it they are a legitimate user; the snippet only discloses a three-layer architecture that separates what from who, not concrete attack steps or implementation details.

Why it matters: HKR-H/K/R all pass: the hook is strong, the post offers a what/who three-layer design, and agent permissions are a live security worry. No real incident, success rate, or product comparison keeps it at the featured threshold.