Review · August 21, 2026
Signed Authorization-Evidence Records for WIMSE-Authorized AI Agent Actions
Problem Statement
This draft addresses a specific audit gap in the companion AI agent authentication and authorization draft (by Kasselman et al.): that draft specifies how AI agents obtain identity and acquire runtime authorization, but it does not define a durable signed record of the authorization decision itself. This spec proposes a "Permit" — a COSE_Sign1-signed record capturing the pre-execution authorization decision — and defines how WIMSE/SPIFFE identities, OAuth tokens, HTTP Message Signatures, and Shared Signals Framework events integrate with it.
How the Spec Fits In
This is a companion profile to two other individual Internet-Drafts, neither of which is a working group document:
draft-munoz-scitt-permit-profile— defines the actual Permit object (COSE_Sign1 envelope, Receipt construction, verification rules). This document defers the core data structure to that draft.- The Kasselman et al. AI agent auth draft — whose Section 11 audit requirements this document maps to.
This spec does not define new protocols or token formats. It provides WIMSE-specific integration guidance layered on top of those two companion drafts, plus SCITT, SPIFFE, HTTP Message Signatures, OAuth 2.0, Transaction Tokens (RFC 8693), and Shared Signals Framework. The dependency chain is thus: this spec → SCITT permit profile (individual draft) → companion auth draft (individual draft) → WIMSE, SCITT, OAuth, HTTP Message Signatures, SPIFFE. That's a deep stack of unratified individual drafts at the base.
High-Level Summary
When a WIMSE-authorized AI agent takes an action, the authorization decision is currently ephemeral — there is no tamper-evident record that the decision was made, by whom, and with what scope. The Permit fills that role: a signed, long-lived record that commits cryptographically to the canonical request bytes before dispatch, and (via a paired "Closure Record") to the actual dispatched request.
This draft specifies the WIMSE flavor of a Permit:
- Identity binding: SPIFFE URIs or WIMSE workload identifiers are typed as
subject_type+subject_idfields in the Permit. - Audit crosswalk: Maps the eight audit requirements from the companion draft's Section 11 to specific Permit fields (e.g.,
subject_idfor authenticated agent identifier,idempotency_keyfor correlation). - Optional OAuth integration: An authorization server may embed a private
permit_idclaim in an access token, allowing a resource server to retrieve and verify the corresponding Permit. The claim name is unregistered. - Optional HTTP Message Signature integration: Header coverage beyond the request body may be added via a recorded HTTP Message Signature in the Closure Record.
- Optional Transaction Token linkage: A Transaction Token's transaction ID may be placed in
idempotency_keyandcorrelation_idto correlate runtime context with durable evidence. - Optional SSF/CAEP/RISC eventing bridge: A descriptive (non-normative) pattern for propagating Permit lifecycle state transitions as real-time SSF events.
- Authorization lineage for delegated actions: Describes how parent/child Permit relationships enable verification that a child agent's authority is equal to or narrower than its parent's.
Interoperability Assessment
Significant Interoperability Concerns
The majority of integration mechanisms in this draft are explicitly marked optional, and the reference implementation (Keel API's open-source Issuer) does not implement most of them: delegated-subject fields, HTTP Message Signature recording, OAuth permit_id claim, Transaction Token linkage, and the SSF bridge are all described as future work. What is actually implemented — SPIFFE subject typing and the Section 11 field mapping — is straightforward and could reasonably be documented in a README rather than an Internet-Draft.
The permit_id claim name for OAuth tokens is explicitly noted as unregistered. Any deployment adopting this pattern will choose its own claim name, defeating the interoperability goal for that mechanism.
The Permit object itself — the COSE_Sign1 envelope, field definitions, receipt construction — is entirely delegated to draft-munoz-scitt-permit-profile, another individual draft with no working group standing. This means two individual drafts must advance in lockstep for either to be implementable from standards text alone. Reviewing or implementing this document in isolation is not fully possible.
The "Authority Lineage" verification mechanism (parent/child Permit relationships, Comparator Profiles, Authority Attenuation rules) is explicitly out of scope: "This WIMSE profile does not define those formats." It is unclear what normative value the Authorization Lineage section (3.6) then provides beyond aspirational description.
Mechanisms That Could Be Removed
Sections 3.3 (HTTP Message Signature Integration), 3.4 (OAuth Token Composition), 3.5 (Transaction Token Linkage), and 3.7 (SSF Eventing Bridge) are all non-normative patterns that the reference implementation does not implement and which are already composable via existing specifications without this document. Removing these sections would reduce the spec to its implementable core: SPIFFE identity typing and the Section 11 audit crosswalk, which is genuinely useful but narrow in scope.
Writing Quality
The writing is competent and the document structure is clear. RFC 2119 key words are used appropriately. The scope section (1.1) is unusually explicit about what the spec does and does not cover, which is a good practice.
The draft is moderately dense with cross-references and acronyms (Permit, Closure Record, WIT, WPT, Authority Lineage, Comparator Profile, Authority Attenuation, Authority Representation) most of which are defined only in other individual drafts. A reader without the companion SCITT profile draft open alongside will find many terms undefined in context.
The "Implementation Status" section (Section 8) is refreshingly honest — it clearly flags that most of the interesting composition patterns are not implemented — but this transparency also underscores that the spec is largely aspirational at this stage.
Section 3.6 on Authorization Lineage reads more like a requirements statement for a future profile than normative text. It would benefit from either being removed or clearly marked as informative.
Author Background
Christian Munoz ([email protected], Keel API, Inc.)
- RFCs published: 0
- Active Internet-Drafts: 2 (
draft-munoz-wimse-authorization-evidenceanddraft-munoz-scitt-permit-profile) - Expired drafts: None
Christian is a new IETF participant with no published RFCs and only two active drafts, both of which are closely related (this spec and its companion SCITT profile). There is no prior IETF engagement visible in the Datatracker profile. The drafts appear to be written to document a product (Keel API) rather than to solve a working-group-identified interoperability problem, which is consistent with the "individual draft" status and the gap between what is specified and what is implemented.
Overall Assessment
This draft describes a genuinely real gap — the absence of a durable audit record for AI agent authorization decisions — but the solution as written is premature for IETF standardization. The core data structure lives in another unratified individual draft; most of the integration mechanisms are optional and unimplemented; and the dependency chain bottoms out in two individual drafts rather than working group documents. The spec would benefit substantially from: (1) consolidation with the companion SCITT permit profile, (2) removal of the optional/unimplemented composition sections, and (3) engagement with the WIMSE and SCITT working groups to determine if there is consensus interest in the problem before investing further in the solution.