Update review · November 21, 2026
Signed Authorization-Evidence Records for AI Agent Actions Using WIMSE Workload Identity (Update)
What Changed
Revision 02 is a correction pass. It adds no new mechanisms and removes none.
The draft still has the same seven sections of integration guidance. Most of the edits make the text claim less. Several statements in 01 were wrong, and the new change log says so directly.
The biggest structural change is outside the author's control. The parent document was adopted by the WIMSE working group in September 2026 as AI Identity Management System (AIMS). This draft now references draft-ietf-wimse-aims-01 instead of the individual draft draft-klrc-aiagent-auth.
Was the Previous Feedback Addressed?
Mostly no. One concern was resolved by the AIMS adoption, and one got worse.
The dependency chain: partly resolved
The previous review said the stack bottomed out in two individual drafts. One of them is now a working group document.
The other is not. The Permit object still lives entirely in draft-munoz-scitt-permit-profile, an individual draft by the same author. The two drafts were not consolidated. You still can't implement this document without the companion open next to it.
Optional, unimplemented sections: not addressed
The previous review suggested removing Sections 3.3 (HTTP Message Signatures), 3.4 (OAuth access token claim), 3.5 (Transaction Tokens), and 3.7 (SSF bridge). All four remain. All four are still unimplemented in the reference Issuer.
The implemented core shrank
This is the most important change in the revision.
The previous review identified SPIFFE subject typing and the Section 11 crosswalk as the implementable core. Revision 02 walks both back.
- SPIFFE subject typing is no longer implemented. The Implementation Status section now says the Issuer "does not implement the SPIFFE or wimse-scheme subject typing in Section 3.1." Its signed subject types are
agent,user,service_principalandsystem. - The typing rules dropped from MUST to SHOULD. Setting
subject_typetospiffeorwimseis now a recommendation. Open Issue 5 asks whether to make it a requirement again. - The crosswalk no longer maps requirements to fields. The table column changed from "Permit Field" to "Evidence available or needed." Rows for posture, remediation, and correlation now describe evidence the deployment has to supply itself.
The draft now states plainly that "the decision record covers only part of the AIMS audit minimum." That is more accurate than 01. It also means the one normative requirement that tied this profile to WIMSE identity is gone.
So what does a WIMSE profile of Permits require now? Very little. The remaining MUSTs say that a Permit reference can't substitute for verifying the Permit.
The reference implementation is not open source
Revision 01 said the reference Issuer was published in the keel-api repository under Apache 2.0. That was wrong. Revision 02 corrects it: keel-api is a proprietary hosted service and its source is not public.
The open artifacts are the Permit specification, JSON schemas, test vectors, a verifier package, and client SDKs. The previous review's concern that the draft documents a product stands, and is now stated by the draft itself.
The unregistered claim name: not addressed
The illustrative OAuth claim was renamed from permit_id to permit_ref. It is still unregistered and still described as an issuer-private claim. Two deployments will still pick two different names.
There is one real improvement here. The claim value can now carry a digest of the Permit alongside the identifier, so a resource server can detect a substituted record. This follows the reference projection in draft-liu-oauth-authorization-evidence.
Authorization Lineage: not addressed
Section 3.6 is essentially unchanged. It still defers every format to the companion SCITT profile and still reads as a requirements statement for future work.
Corrections Worth Noting
Several fixes improve the draft's accuracy.
- "WIMSE-authorized" is gone, including from the title. WIMSE provides workload identity and authentication. Authorization in AIMS uses OAuth. The old term was a category error.
- Transaction Tokens are described correctly. Revision 01 suggested putting a transaction identifier in
idempotency_key. That breaks when several actions share one transaction. The draft now says each action needs its own idempotency key, and correlation goes in the Closure Record'scorrelation_id. - HTTP Message Signatures now follow the WIMSE profile. The section references draft-ietf-wimse-http-signature for component coverage. The Issuer no longer computes its own signature over the request. The deployment preserves the workload's signature instead.
- The "two independent commitments" claim was removed. The security considerations now say that neither the HTTP signature nor the body commitment "establishes which bytes reached the provider." That is an honest limit on what this evidence proves.
- SCITT terminology is tighter. The draft now separates a COSE Receipt from the companion's "provisional linked-chain proof." A verifier claiming SCITT registration must verify a real COSE Receipt.
- AIMS is quoted accurately. Revision 01 said the parent draft was explicit that the record format is left to implementations. It says nothing about format.
What's New
The Background section now surveys competing work. It names draft-liu-oauth-authorization-evidence, draft-mcguinness-oauth-mission, draft-chueayen-attestation-receipts, draft-pidlisnyi-aps, and draft-thaha-wimse-agent-evidence-binding.
That last one matters. The draft says it "maps the same AIMS audit fields to execution receipts." There are now at least two individual drafts in WIMSE proposing an evidence format for the same AIMS Section 11 requirements. The draft positions itself against the others in one sentence and does not argue why a Permit is the better fit.
A terminology note separates this draft's "authorization evidence" from the authorization_evidence type in the Liu draft. The name collision is real and will confuse readers of both.
The draft also adds a pointer to OpenID AuthZEN, which AIMS identifies as the interface to an external policy decision point.
A change log was added as Appendix B. It is detailed and candid about what 01 got wrong.
Author Background
No new authors. Christian Munoz of Keel API remains the sole author.
Overall Assessment
Revision 02 is a more accurate document than 01. It is also a thinner one.
The corrections removed overclaims, and what's left is a set of optional composition patterns around a Permit format defined elsewhere. None of the WIMSE-specific parts are implemented, including the subject typing that 01 presented as working.
The AIMS adoption gives this draft a real anchor. AIMS Section 11 requires tamper-evident audit records and defines no format. That gap is now a working group problem, and several drafts are competing to fill it.
The useful next step is the same as before, with one addition. Consolidate with the SCITT permit profile, cut the unimplemented sections, and take the Section 11 format question to the WIMSE list alongside the other evidence drafts. The working group should decide whether it wants one evidence format before reviewing four.