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

Update review · November 16, 2026

Cross-Org Delegation -02, R10 Takes Over the Document

draft-reece-wimse-cross-org-delegation -02View on Datatracker →

What changed in -02?

The document went from 11 pages to 17. Almost all of the growth is about one requirement.

The -01 review covered a problem statement with seven facets and ten requirements. The facets and the count of requirements are the same in -02. R1 and R3 through R9 are word-for-word unchanged.

R10, the execution-time human approval requirement, roughly doubled in length. It is now about 27 lines. No other requirement is longer than eight.

What was added

  • Acquisition versus binding (Section 3.2). Getting trust material for an unknown organization is one problem. Knowing the material belongs to the real-world organization is another. R2 now says it only demands the first.
  • First-contact residual (Section 7). No mechanism removes the first leap of trust. The draft says the cost is paid once.
  • Consume only on admission (R10, Section 5). An approval is used up only when the action is admitted. Evaluation and refusal MUST NOT consume it. A reservation MUST be released on refusal.
  • Canonicalization or mapping profile (R10). If the approval evidence and the action use different representations, a declared profile MUST define the match. A missing or ambiguous mapping is a failed verification.
  • Sole-authority assumption (Section 6). The requirements assume the delegated authority is the only path the agent has. If the agent also holds the user's own credentials or cached sessions, R1 through R10 say nothing about whether the delegation governed the action.
  • Implementation Status (Section 8). Two outside reports, EMILIA Protocol and OASNT, plus a one-line claim from the author.
  • Three references. RFC 8693 is now cited in Section 4, plus two individual drafts.

Does the new part hang together?

Partly. The new material is careful, and some of it is useful. The problem is where it sits.

The approval-burning attack is a real attack vector. Section 7 describes a relying party that consumes an approval during evaluation and then refuses on a mismatch. An attacker can present a near miss on purpose and burn the user's approval. The user approved a payment, nothing happened, and the logs show correct fail-closed behavior. That is worth writing down.

It is also solution design. Reservations, release on refusal, and two attempt records per replay are implementation guidance. Section 1 says the document does not propose a mechanism.

"Reliance unit" is never defined. R10 and Section 7 both depend on it. Section 2 has no entry for it.

Section 7 now names a mechanism. The first-contact text says "a hash-chained authority history and continuity signatures" carry the binding across key rotation. That is a specific design inside a document that calls itself mechanism-neutral.

Section 7 also adds normative requirements outside the requirements list. There is a new MUST NOT on blind retry and two new SHOULDs on separate recording. A solution author reading only Section 5 will miss them.

The Implementation Status section reports a failure. Both outside implementations cover only R10. Neither carries the human-approval designation inside the delegated authority. OASNT keeps it in relying-party route policy and is marked "NOT MET and not currently attempted." That clause is the one part of R10 that is about cross-organizational delegation. The draft concludes that current formats cannot express it. Another reading is that the clause is asking for too much.

The author's own implementation claims all ten requirements. Section 8 gives no details and defers them to "mechanism-track documents." None are cited.

Was the previous feedback addressed?

Partly addressed:

  • R10 is under-specified. The at-most-once behavior is now much more precise. "Execution-time human authorization" and "designated" still have no definitions in Section 2.
  • Engagement with existing work. RFC 8693 is cited. Section 3.7 still refers to OAuth transaction-authorization proposals without citing any of them. There is still no reference to Rich Authorization Requests or Transaction Tokens. The note on page 1 still promises those references in a later revision.

Not addressed:

  • R10 as a MUST. I suggested a SHOULD until the replicated-verifier problem has an answer. R10 has more MUSTs now. The sentence about distributed consumption in Section 7 is unchanged.
  • The Non-Goals tension. Section 6 still says the document does not select credential formats or schemes. R10 now requires a declared canonicalization profile. The gap between the two got wider.
  • Dropping or splitting R8. R8 is unchanged. Section 7 now adds recording guidance that "composes with R8," so audit got bigger.

Mechanisms added or removed

Nothing was removed. No protocol mechanisms were added, because this document has none. What it added is constraints on how a future solution handles approvals.

Background of the authors

No authors were added. Morgan Reece is still the sole author.

A new Acknowledgments section credits Iman Schrock, Barak Shelef, Songbo Bu, and Arun Thallapelly. The change log attributes most of the R10 text to their review.

Questions worth raising

  1. Should R10 be its own document? It is now most of the diff, all of the implementation reports, and most of the security considerations. R1 through R9 got no attention in this revision.
  2. Is in-chain designation carriage a requirement or a wish? Two implementations skipped it. If current WIMSE and OAuth token formats cannot carry it, the working group should hear that as the main finding.
  3. Does the sole-authority assumption hold for real agents? Most agents deployed today run with the user's own credentials. Section 6 puts that case out of scope. That may put most of today's deployments out of scope too.