Attestation
A Paper Proposes Binding Agent Identity to Zero-Knowledge Proofs, Not Just Keys
Most agent identity schemes stop at key-based authentication: an agent presents a signature, and that's treated as proof enough. A new paper from researchers including Dacheng Tao argues this is the wrong foundation entirely. The core problem they identify: traditional key-based authentication guarantees neither the operator's physical identity nor the agent's code integrity. An agent can hold a valid key while running substituted or tampered code, and no one downstream would know.
Their proposal, BAID (Binding Agent ID), tries to close that gap by combining three things rather than one: local biometric binding, decentralized on-chain identity management, and verifiable agent authentication. The goal is a framework that provides a unified solution for identity verification, trusted discovery, and accountability tracking, ensuring the agent running is the one the binding user actually authorized, executing the exact code that was committed.
The technical mechanism is a zkVM-based authentication protocol. Instead of just checking a signature, the system uses recursive zero-knowledge proofs to cryptographically bind the agent's execution to its registered identity and authorized user. The stated purpose is preventing code substitution attacks and ensuring execution integrity without revealing sensitive operational details — meaning you can prove 'this exact code ran, operated by this exact binding' without exposing the code or the operator's identity to every party that checks.
The authors built a working prototype rather than stopping at the proposal stage. They developed the full system architecture including identity management smart contracts and zkVM verification circuits, and report gas-efficient on-chain operations and millisecond-level verification latency suitable for real-world deployment. That latency number matters: zero-knowledge proof systems have historically been too slow for anything resembling live authentication, so a millisecond-scale result — if it holds up under independent testing — would be the difference between an interesting cryptographic exercise and something deployable.
What's notable is where this sits relative to the rest of the agent-identity field. Other recent proposals — biometric-bound delegation tokens, OTEL-based audit trails with agent claims, China's split-knowledge national identity layer — all wrestle with the same tension: accountability usually gets purchased by making agents identifiable to everyone they interact with. BAID's zkVM approach is an attempt to decouple those two things at the cryptographic layer instead of the institutional one, proving code integrity and user binding without handing over the underlying secret. Whether that proof style scales past a prototype, across multi-agent chains and real production workloads, is the open question the paper doesn't yet answer.