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

Review · November 21, 2026

Evidence Binding for AI Agent Credentials

draft-thaha-wimse-agent-evidence-binding -03View on Datatracker →

The Problem

AIMS says agent audit logs must be tamper-evident and lists seven fields every audit event must record. It deliberately doesn't pick a log format. It also doesn't say how someone holding an agent's credential finds that agent's logs.

This draft fills that gap with the authors' own SCITT-based evidence format. It adds a token claim that points from an agent's credential to its evidence. It then goes further and defines how a service decides, on every call, whether the calling agent is currently "assured" enough to be let in.

How It Fits In

This is a companion to AIMS, and it says so. It makes no change to AIMS.

The real dependency is Agent Execution Evidence, an individual SCITT draft by the same two authors. Seventeen terms are imported from it, including Agent Registration Authority, Assurance Statement, Evidence Service, Serving Party and Supervisory Authority. None of these roles exist in WIMSE or OAuth today.

It touches a lot of other work:

  • A new claim for the WIMSE Workload Identity Token, or any JWT or CWT that identifies an agent
  • Nested act claims from Token Exchange for the actor chain
  • The txn identifier from Transaction Tokens
  • SCITT receipts and transparency services (RFC 9942, RFC 9943)
  • An in-toto predicate to sit alongside the Gilda agent audit record draft

Does this belong in WIMSE? Only partly. The claim in the workload credential is a WIMSE topic. Audit formats, assurance schemes and registration authorities are closer to SCITT and the proposed audit work.

What It Defines

The draft says it proposes "one claim, one predicate type and one HTTP field". That undersells it.

The aee_evidence claim. The identity server puts this in the agent's token. It has ten members. Three are required: the Evidence Service identifier, a digest of the governing policy pack, and a logical_agent_id. The logical ID is a did:web identifier issued by an Agent Registration Authority. It names the agent across every runtime it's deployed in, where the token's sub only names one workload.

Optional members carry digests of the agent's credential statement, registration statement and Assurance Statement. The token can also carry a cached assurance level and status, with a status_checked_at timestamp saying when the identity server last looked.

After-the-fact verification. A relying party gets a Bundle of evidence, verifies it against trust anchors it already has, then checks that the issuer, runtime identity, logical identity and policy digest all match the token.

Per-call admission. A Serving Party runs seven steps before acting on any call. It validates the token, checks registration and assurance status, checks the Principal and the current actor, verifies the caller's digest of the request, and writes an admission receipt whether it admits or refuses.

The AEE-Caller-Ref header. The calling agent writes a receipt before sending the request, then sends a pointer to it: Evidence Service, sequence number, receipt hash, transaction ID and a digest of the call. The Serving Party recomputes the digest itself and refuses on mismatch.

An in-toto predicate type that wraps an Execution Receipt so one toolchain can carry it next to the Gilda audit record.

Interoperability Concerns

It can't be implemented on its own. The receipt format, call_digest computation, status values, assurance levels, Bundle verification and Linked Verification all live in the companion SCITT draft. That draft is an individual submission at -02. This draft only makes sense if both advance together.

The pieces a relying party needs most are out of scope. Bundle discovery and retrieval are "supplied by deployment configuration or another protocol". Trust anchors for the Evidence Service and Transparency Service are obtained out of band. So the claim tells you who should have the evidence, and nothing about how to get it.

The admission decision rests on undefined inputs. assurance_level is an integer whose meaning comes from an assessor and scheme the Serving Party "accepts". RESTRICTED status allows operations "within the restricted class the scheme defines". Two Serving Parties will reach different answers for the same token.

The cached status couples the identity server to the assurance ecosystem. The WIT issuer now has to poll an assurance status list before issuing tokens. That's a new job for a workload identity issuer.

AEE-Caller-Ref has no integrity protection. The draft never says the header must be covered by the WIMSE proof token or HTTP message signature. It also uses no normative keyword for sending it: the calling agent just "includes" it.

Replay detection requires unbounded state. A Serving Party must refuse any caller reference it has already admitted. No expiry window is given.

principal_link is a hash of the access token in the Authorization header. That breaks when a gateway exchanges or re-issues the token between the two parties.

The sub mapping is left open. If sub names a user instead of the workload, the mapping to the agent's runtime identity "MUST be separately specified". That covers most OAuth access tokens issued to agents today.

What Could Be Removed

  • The AIMS field mapping in Section 3. It's informative and the draft itself calls the mapping "partial". It belongs in an appendix of the SCITT draft.
  • The in-toto predicate in Section 7. It's a format bridge between two individual drafts, and neither needs it.
  • The cached assurance_level, assurance_status and status_checked_at. They're an optimization with three paragraphs of freshness rules, and high-risk operations are told to check the status list anyway.
  • Most of the admission procedure. Steps 2 through 4 are relying-party policy.

What's left is a small spec: a claim with an Evidence Service, a policy digest and a logical agent ID, plus a header that links one call to one receipt. That version would be much easier for WIMSE to evaluate.

Writing Quality

The writing is good. It uses BCP 14 language correctly, registers its claims and header properly, and reads like an IETF document.

It's also unusually honest about its limits. "The claim is a pointer and a commitment, not a proof." A receipt shows a statement was registered, not that it's true. The actor chain is "attribution, not authority". Showing that authority narrowed at each hop is listed as an open issue.

The problem is density. The draft says it's "short by design" and runs 26 pages. Nearly every sentence points into a section of the companion draft, so you can't review this one without the other open. Many paragraphs are lists of caveats on the mechanism they just defined.

The change log shows a fast-moving document. This is the fourth revision, and -01 was uploaded from the wrong commit. Named reviewers are credited for specific fixes, so the authors are getting and acting on feedback.

Authors

Shameer Thaha has no published RFCs. Their Datatracker profile lists two active drafts: this one and the companion SCITT draft. No expired or replaced drafts.

Shaibi Shamsudeen has the same record: no RFCs and the same two active drafts.

Both are new IETF authors, listed as individual contributors with no affiliation. All of their IETF output is this one proposal split across two working groups.

Bottom Line

So what should WIMSE do with this? The useful question it raises is real: given an agent's credential, where is its audit evidence? A single claim answering that could be a reasonable WIMSE extension.

The rest is an assurance and registration ecosystem with new authorities that no working group has agreed to. That discussion belongs with the SCITT draft first, since this document can't stand without it.