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

Review · October 26, 2026

Actor Chain -02, Six Profiles and a Bootstrap Endpoint for Multi-Hop Delegation

draft-mw-oauth-actor-chain -02View on Datatracker →

The problem it's solving

A request passes through services A, B, and C before reaching resource D. Token Exchange lets each hop add itself to a nested act claim, but RFC 8693 says earlier actors in that chain are informational only. There are no rules for how the chain is built, what each hop is allowed to see, or how D can check that B actually participated rather than the AS just saying so.

This draft adds those rules. Each verified hop produces two signed artifacts: the actor signs a proof of what it asked for, and the AS signs an acceptance of that exact proof. Acceptances chain by hash. A resource can require the AS's word alone, the current actor's signature, signatures for a disclosed subset of actors, or the full signed path.

How it fits in the working group

This is a profile of Token Exchange with a large footprint. It defines a new AS endpoint, four JWT claims (actp, acti, actc, acte), five token request parameters, two AS metadata fields, one Protected Resource Metadata field, four introspection members, a token type URN, and three media types. It normatively depends on Resource Indicators, Protected Resource Metadata, AS Metadata, Introspection, DPoP, JCS, and the still-unfinished Identity Chaining WG draft for cross-domain handoff.

It is one of at least five individual drafts aimed at the same question. The McGuinness Actor Profile, Receipts, and Proofs suite puts actor-signed proofs inline. Delegation Chain carries AS-signed hop records. The Fletcher transaction token chaining profile and the WIMSE agent delegation chain draft overlap too. This one is the oldest of the group, with lineage back to a SPICE draft in February 2026, and the abstract and Section 29.2 go out of their way to say so.

What the mechanism actually does

Six profiles, picked once per workflow. Assurance is Declared (AS says so) or Verified (actor signs too). Disclosure is Full, Subset, or Actor-Only. The profile goes in actp and can't change for the life of the workflow identified by acti.

Starting a chain takes two round trips. The first actor calls a new actor_chain_bootstrap_endpoint, authenticates as an OAuth client, and gets back an opaque context plus the values it is about to sign: its own actor ID, the workflow identifier, a workflow subject the AS allocates, the accepting issuer, and the target resource and maximum scope. It then signs a root proof over those values and redeems the context at the token endpoint with grant_type=client_credentials. Declared profiles use the same bootstrap without the proof.

Each later hop is a Token Exchange with three extra parameters. The actor sends the inbound token as subject_token, plus actor_chain_profile, actor_chain_step_proof, and actor_chain_view_opening. The proof is a JWS the actor signs over its identity, the workflow, the accepting AS, the exact resource and canonical scope, a fresh request_id, and a hash of the previous acceptance.

The actor commits to what it knows without revealing it. The proof carries a salted SHA-256 over the JCS of the actor list the actor was allowed to see. The actor sends the opening (salt plus list) to the AS, which checks it. The opening never appears in the token unless the resource asked for full path evidence.

The AS signs an acceptance. actc is an AS-signed JWS over the hash of the exact proof bytes, chained to the previous acceptance by hash. Every Verified token carries the latest one. This is the chain head. Re-encoding breaks it on purpose.

Evidence is optional and chosen by the resource. The resource advertises actor_chain_evidence_required in its metadata with a coverage and allowed delivery modes. Coverage current is the latest proof and acceptance. path is every proof, acceptance, and opening from root to now, and only works with Full disclosure. disclosed, new in -02, is the original proof and acceptance for each actor that happens to be in act, with the issuer vouching that those records belong to this path. Delivery is inline in acte or fetched by introspection.

Cross-domain handoff is a preserve, not a hop. The source AS uses Identity Chaining with a new token type to move the protected state to the destination AS byte for byte. The destination AS adds no acceptance. The next actor's proof hashes the foreign acceptance directly.

Replay is a distributed-systems requirement. The AS must atomically reserve the tuple of issuer, actor, workflow, and request_id in durable, linearizable storage, fence stale workers after crash takeover, return the identical acceptance on exact retry, and return 503 rather than guess when it can't reach that storage.

So what is new in -02?

The document was split. The 109-page -01 became a 63-page protocol spec plus a short Informational overview draft. Normative keyword count dropped from about 435 to about 227. Suresh Krishnan joined as fifth author. All author affiliations were removed from the header. The disclosed coverage mode and envelope version 2 were added. The title changed from "Cryptographically Verifiable Actor Chains" to "Protocol and Interoperability".

Significant interoperability concerns

The combinations don't converge. Six profiles, four evidence postures, two delivery modes, and two envelope versions. Section 6 says an implementation MAY support any subset. Two conformant AS deployments and two conformant resources can have zero overlap. Capability metadata lets them discover that, but discovery is not interoperability.

There is no way to start a chain from a user. Root is client_credentials only and the AS allocates the workflow subject. The draft says the subject "identifies the party or context represented by the workflow," but nothing describes how a human's authorization code grant becomes the first link. For the agent use cases in Section 35, where a user delegates to an orchestrator, the user never enters the protocol.

Actor key trust is out of scope. Every verification step says the key must be "independently authorized for the exact ActorID" at event time, with historical key context retained for as long as any artifact might be checked. How a resource learns which key belongs to https://actors.example/b is never specified. This is the same gap the Actor Proofs draft has, and it is the whole security model.

Envelope versions are feature flags. Version 1 means current or path. Version 2 means disclosed. A version 2 parser must reject version 1 coverage values and vice versa. Resources advertise a version and coverage pair. That's a version number doing the job of an enum.

The empty-scope rule will break on real servers. Section 8 requires that when scope is omitted from a Verified transition, the AS's configured default scope MUST be the empty set, or the request MUST fail with invalid_scope. Most authorization servers have non-empty defaults. Every Verified deployment will have to always send scope explicitly to avoid this, which makes the rule dead weight.

Declared profiles add almost nothing over RFC 8693. They add actp and acti to the token and require the bootstrap endpoint. The AS was already free to populate act however it liked. Three of the six profiles exist to have a lower tier, not to solve a problem.

Path evidence grows quadratically. Each opening repeats the full actor list up to that point. The draft measures a depth-10 inline token at roughly 24 KB. Nobody is putting that in an Authorization header, so path evidence is introspection-only in practice, which the draft doesn't say.

Mechanisms that could be removed

  • The bootstrap endpoint. A new AS endpoint and a two-round-trip root flow exist so the actor can learn the workflow identifier before signing. Let the AS mint acti on first exchange, or let the root actor pick it. Either removes an endpoint, two parameters, and a page of error handling.
  • Declared profiles. Fold acti into a one-paragraph claim definition and drop the rest.
  • disclosed coverage and envelope version 2. It verifies signatures but trusts the issuer for membership and order, which is most of what Declared already gave you. Added in -02 and not yet tested.
  • Introspection delivery. Defer until inline sizes force it. It brings a parallel registry, a parallel validation path, and new availability failure modes.
  • Subset narrowing during preserve. A page of rules for an edge case.
  • The replay-store engineering in Section 20. Keep the uniqueness requirement. Move linearizability, fencing, and crash takeover to Security Considerations as deployment guidance.
  • Sections 27 through 35. Worked examples, deployment choices, rationale, test vector notes, checklist, future work, and the MCP, A2A, and Kubernetes tour. That's 10 pages, all informative, and an overview draft already exists to hold them.

What remains is the proof, the acceptance, the commitment, the exchange parameters, current and path evidence, resource metadata, and cross-domain preserve. That's a 30-page document.

Writing quality

Much better than -01, and still exhausting. Nearly every sentence is a rule. Many rules are about what a party MUST NOT "describe" or "treat as," which is policy framing dressed as protocol. Every internal cross-reference reads "Section Section 17," a tooling bug on every page. Examples use placeholders like <proof_A> and the test vectors are "retained locally by the authors," so nothing in the document can be checked byte for byte. The abstract spends its last two sentences on draft lineage, and the lineage appears again in Sections 1.4 and 29.2. That reads as a priority claim against the competing drafts rather than information a reader needs.

Background of the authors

Five authors on -02, up from four. Datatracker profiles were not re-fetched for this post. Figures below are from the August review of this group's WIMSE transitive attestation draft, and all affiliations were dropped from the header in this revision.

  • A. Prasad (lead contact, previously listed at Oracle): 0 RFCs, around 7 active drafts across OAuth, SPICE, RATS, and WIMSE. A new author writing at a high rate.
  • Ram Krishnan (previously listed at JPMorgan Chase): 6 RFCs going back to 2002 in MPLS and networking, around 7 active drafts, recently shifted to security topics.
  • Diego R. Lopez (Telefonica): 12 RFCs, around 22 active drafts, chairs the GREEN working group. Network operations background rather than identity.
  • Srinivasa Addepalli (Aryaka): 0 RFCs, around 7 active drafts overlapping with Prasad's.
  • Suresh Krishnan (new to -02): a long-time IETF leader with dozens of RFCs, a former Internet Area AD and IAB member. His name on the header is new and does not appear on the companion overview draft.

The same four people author the overview draft, the SPICE predecessor, the TLS session-bound tokens draft, and the WIMSE transitive attestation draft. This is a coordinated push across several working groups.

Questions worth raising

  1. Which of the five delegation-evidence drafts should the WG compare first? This one is the oldest and the largest. Adopting it would set the shape for the others.
  2. Where does the user go? If the answer is "a future companion," the agent use cases in Section 35 aren't served by this document.
  3. Can the bootstrap endpoint go away? Ask the authors what breaks if acti is minted on the first exchange.
  4. Is anyone implementing this? Two revisions and an overview with no public test vectors or named implementation is a signal for a spec this size.

For the competing inline-proof design, read Actor Proofs. For the cross-domain handoff this draft builds on, read Identity Chaining.