Agent Identity Assurance Framework
Draft 0.1 · not a ratified standardA proposed open standard for who is accountable for an AI agent, how strongly that is evidenced, and what a website can verify.
What identity can prove, and what it cannot.
| Claim | Provable | How |
|---|---|---|
| The request came from the holder of this key | Yes | Signature on the request |
| The key belongs to this agent | Yes | Registry-signed credential |
| The operator is a real legal entity | Yes, at OAL 2+ | Business-registry evidence, human review |
| The identity is valid right now | Within the revocation window | Signed revocation feed, short credentials |
| The agent runs a given model | No | Self-asserted; unverifiable over the wire |
| The agent has not been prompt-injected | No | A compromised agent still holds valid credentials |
| Unregistered traffic is malicious | No | No credential means unidentified, nothing more |
Three axes, graded separately.
Operator (OAL)
Who is accountable?
- 1 · Domain-verified
- 2 · Verified legal entity
- 3 · Accountable entity
Key (KAL)
How well is the key protected?
- 1 · Possession proven
- 2 · Attested non-exportable
- 3 · Attested runtime
Delegation (DAL)
By what authority does it act?
- 0 · No delegation
- 2 · Evidenced delegation
- 3 · Constrained delegation
AgAL = min(OAL, KAL)
The composite level is the weaker of operator and key assurance. Delegation is kept separate.
From registration to a verified request.
Register
Only the public key is submitted.
Prove possession
The agent signs a single-use challenge.
Present
A credential of at most 15 minutes, and a signature.
Verify offline
No call to the registry per request.
GET /api/products HTTP/1.1
Host: shop.example
Signature-Input: sig1=("@method" "@authority" "@path");
created=1790243730;keyid="aid:v1:acme-corp:01J8Z…";alg="ed25519";tag="aidr"
Signature: sig1=:MEUCIQD…:
Agent-Credential: eyJhbGciOiJFZERTQSIsInR5cCI6…
What a conforming website must check.
For website owners: implementing AIAF costs nothing. The standard is Apache 2.0, and verifying needs no account and no call to any registry.
The credential's signature, against a registry on your trust list.
The request's signature, against the key inside the credential.
That the signature covers method, authority and path.
Revocation, treating "could not determine" as distinct from "not revoked".
That declared values are never treated as verified.
That a request without a credential is unidentified, not malicious.
Require identity where it matters. An agent that will not identify itself gets what you give an anonymous visitor, or nothing.
Downloads.
Apache License 2.0: free to implement, including by competing registries.
This draft is edited by the team behind AIDR, a commercial registry. Its text names no vendor, any registry may issue identifiers under it, and it is meant to pass to a neutral standards body.