Attestation
A Zero-Knowledge Proof That the Agent Running Is the Agent You Authorized
A new paper proposes solving a narrower but under-addressed problem: not just who owns an agent, but whether the code actually executing matches what was registered. The BAID (Binding Agent ID) framework combines biometric binding, on-chain identity, and a zkVM-based authentication protocol that cryptographically ties an agent's live execution to its registered identity and code.
The core move is using recursive zero-knowledge proofs so a verifier can confirm the agent running right now is the exact authorized code, without exposing the operational details of what that code does. The paper frames the threat plainly: traditional key-based authentication confirms neither the operator's physical identity nor the agent's code integrity, leaving room for code substitution attacks — an attacker swaps in different logic behind the same credentials while everything upstream still looks authorized.
Most agent-identity proposals this year have focused on binding a credential to a human or an organization. BAID's contribution is the layer underneath that: binding the credential to the specific executable, proven in zero-knowledge, so identity verification and code-integrity verification happen in the same cryptographic step. The authors built a working prototype — identity management smart contracts plus zkVM verification circuits — and report gas-efficient on-chain operations with millisecond-level verification latency, which they argue makes the approach viable for production use rather than just a theoretical construct.
This matters because most current accountability proposals assume the hard part is proving who authorized an agent, then trusting that the deployed code matches what was approved. BAID's framing is that this trust gap — between "authorized identity" and "actual running code" — is exactly where code substitution attacks live, and no existing interoperability protocol addresses it. Whether zkVM proof generation stays cheap enough at scale for real deployments, and whether registries adopt a standard for presenting these proofs during discovery, remains open. But the paper is a useful reminder that identity binding and code-integrity binding are separate problems, and most of the industry's current attention is on the former.