The Scope Creep Automated reviews of new AI-related working group submissions

Review · November 10, 2026

Authorization Evidence and Audit Trail for OAuth 2.0 Access Tokens

draft-liu-oauth-authorization-evidence -02View on Datatracker →

When a user approves an agent's request, the AS knows what the user was shown and what they clicked. Nobody else does. The resource server gets a token, and an auditor gets nothing.

This draft has the AS write that decision down as a signed record and attach it to the access token. The record says who approved, what text they were shown, how they confirmed, and when. It stays verifiable after the token is gone.

Revision -02 is close to a rewrite. It grew from 23 pages to 45. The audit_trail sub-object from -01 is gone. New in -02 are a structured disclosure object, an authority_hash binding to the approved limits, a reference form of the record, and a whole withdrawal mechanism that -01 never mentioned.

How it fits in the working group

It's a new Rich Authorization Requests type called authorization_evidence. RAR section 7.1 lets the AS enrich an authorization details object, and this draft defines one type that does that. That's within the charter.

It depends normatively on about ten RFCs: 6749, 7009, 7662, 9396, the JOSE set (7515, 7517, 7518, 7519), and JCS (RFC 8785).

No new roles. The AS gains a new job as a signer of long-lived approval records. The RS gains a verification procedure. The client just asks for the type.

It's one piece of a larger set. The draft describes itself as the "third object" of the Bounded Work Authorization framework, and it pairs with Domain-Specific Authorization Schemas. Both references are informative, but the terminology section borrows its definitions from the framework draft. You need all three open to follow it.

What the mechanism actually does

  1. The client adds {"type": "authorization_evidence"} to its authorization request.
  2. The AS shows the user a consent screen and records the decision: approved, narrowed or declined.
  3. The AS builds an evidence object. It contains the approver, a disclosure (summary, notices, risks, locale), the confirming action such as button_click, and a timestamp.
  4. The AS adds authority_hash, a SHA-256 digest of the other authorization details object that holds the actual limits, like "at most 50 USD."
  5. The AS canonicalizes the object with JCS, signs it as a detached JWS with typ: evidence+jws, and adds a content_hash of the same bytes.
  6. The token carries either the full record or a two-field pointer (id and content_hash). With the pointer, the RS fetches the full record through Token Introspection.
  7. The RS verifies the signature, both hashes, and that the approval happened before the token's iat.

Then there's withdrawal. The AS gives the user a "withdrawal handle," a URL on the AS. A GET shows a confirmation page, and a POST ends the authorization. The AS then revokes tokens and may update a Token Status List or send CAEP events.

Significant interoperability concerns

The IANA registry doesn't exist. Section 11.1 registers the type in the "OAuth Authorization Details Types" registry "established by RFC 9396." RFC 9396 created no such registry. Types are AS-defined.

The RS can't find the object that authority_hash covers. The rule says to find the authorization details object "whose identifier member carries the value of authority_ref." The member name is whatever the other type picked. The draft lists identifier, work_id or "an AS-assigned instance identifier," and its own example uses transaction_id. A generic verifier has to know every bounds type in advance.

The hash breaks when the AS filters. The digest is taken over the object "as the AS will place it in the token response." RFC 9396 lets the AS issue audience-specific tokens with filtered authorization details. Any difference between the token response, the JWT and the introspection response gives the RS a hash mismatch, and the RS is told to reject.

The audit story has no retrieval path. The use case ends with an auditor fetching the record after the token has expired. The only retrieval mechanism defined is introspection, which returns active: false and nothing else for a dead token. Section 9.7 mentions "a dedicated evidence retrieval endpoint" that is never defined. A reference-form pointer becomes unresolvable the moment the token expires.

Lifecycle state has no wire format. The RS learns that an authorization ended by the AS "reporting the token as inactive or by conveying the current authorization state in a separate member." That member has no name. In practice this is plain token revocation.

The withdrawal handle is protected like a secret and then put in the token. Clients MUST treat it as sensitive and keep it out of logs. It's also a signed member of the evidence object, which in the embedded form is readable by every RS that receives the access token. The draft also says a client can use it for "agent-initiated self-withdrawal," but the only defined interaction is a browser page where the approver logs in.

Evidence depends on the client asking. The RS is the party that wants the record, but the client decides whether to request the type. An RS that requires evidence has no way to signal that.

JCS plus detached JWS. Canonicalizing JSON is a known source of subtle mismatches, especially around numbers. The worked example in Appendix B helps. A compact JWS with the record as its payload would avoid canonicalization completely. content_hash would then only be needed in the pointer form.

Mechanisms that could be removed

A lot of this draft describes records the AS keeps for itself. The draft even says so in a few places.

  • declined evidence. No token is issued, and the record "is not reachable through token introspection." It never leaves the AS.
  • ending_reason and ended_at. The IANA section states these "never leave the Authorization Server."
  • proposal_ref. "An opaque URI assigned by the AS for internal correlation," with no way to dereference it.
  • action. Free-form, and the draft says the RS has no basis for acting on it.
  • evidence_ref. One sentence about delegation chains, with no processing rules.
  • content_hash in the embedded form. The draft explains at length that it adds no security when it travels with the record.
  • The withdrawal section. A page on the AS where users revoke access is AS user interface. The URL format, the GET and POST rules, and the entropy requirement don't affect any other party. If clients need to link users to that page, User Authorization Management URI already proposes AS metadata for it.

What's left is a small and useful spec: a signed approval record, a hash tying it to the approved limits, and a pointer form.

Is the signature worth it?

Section 9.4 is candid. The signature proves the AS recorded an approval. It doesn't prove the user approved. The AS runs the consent screen and holds the key.

Inside a JWT access token, the token signature already covers the same bytes. The second signature only pays off when the record is separated from the token, which is the audit case. That's the case with no defined retrieval path.

Writing quality

The grammar is clean and the BCP 14 usage is careful. The security considerations are better than most individual drafts. They state what each mechanism does not prove. The worked signature example with a real test key is the kind of thing implementers ask for.

The spec is too long for what crosses the wire. The limits of content_hash are explained four times. The security considerations defend design choices at essay length, and the abstract argues ("An approval record that cannot be ended is not an accountable record") instead of describing. The field definitions are long paragraphs squeezed into narrow table cells, which makes the core data model the hardest part to read.

The -01 to -02 change roughly doubled the page count, and most of the growth is the withdrawal mechanism and alignment with the bounded work framework. The evidence record itself got simpler.

Background of the Authors

Datatracker was unreachable for this review. The numbers below come from our earlier 2026 reviews.

  • Dapeng Liu (Alibaba): 3 RFCs from 2014 and 2015 in distributed mobility management, around 10 active drafts and 80+ expired ones. An experienced author who is writing a lot right now across OAuth, WIMSE and agent authorization.
  • Hongru Zhu (Alibaba): 0 RFCs, around 6 active drafts, all in OAuth, WIMSE and agent authorization. A newer author, active in this space.
  • Suresh Krishnan (Cisco): dozens of RFCs, former Internet Area AD and IAB member. Recently added to several agent authorization drafts, including the actor chain documents.
  • Aaron Parecki (Okta): 2 RFCs, RFC 9728 (Protected Resource Metadata) and RFC 10017 (OAuth 2.0 for Browser-Based Applications). Around 16 active drafts, SCIM WG chair, and co-author of the JWT Authorization Grant with Interaction Response draft this spec builds on for asynchronous approval.

The author list is unchanged from -01.