AI agents now pause for approval before they send money, delete files, or email outsiders, and that approval is usually treated as a second factor. A new arXiv paper challenges this assumption, arguing that approval only qualifies as a true second factor under specific conditions.
The paper contends that approval is a second factor only when compromising the agent cannot produce it, and when the approval is bound to the specific effect. The title's phrase "Approve what, exactly?" points to the need for effect binding and trusted rendering, so the human sees and authorises the precise action.
Without these safeguards, an attacker who compromises an agent could potentially generate approvals or redirect them to different actions, undermining the security that second-factor authentication is meant to provide. The paper thus highlights a gap between common practice and actual security guarantees in agent authorisation.