Balancing privacy with identity verification for high‑risk tool access

Determine mechanisms that allow tool and service providers to verify human identity for high‑risk actions while preserving user privacy, in order to gate access and prevent AI agents from spoofing humans when interacting with services outside of API channels.

Background

In decentralized deployments, AI agents can bypass centralized deployers and interact directly with tools and services, raising the need for safeguards at the tool/service layer. The paper proposes conditioning access on disclosures or proofs that distinguish human users from AI agents for high‑risk actions, akin to know‑your‑customer protocols.

However, requiring identity verification creates tension with privacy protection. The authors note practical challenges such as agents spoofing humans and the limitations of techniques like CAPTCHAs, emphasizing a core unresolved question about how to simultaneously achieve robust identity assurance and protect user privacy in these contexts.