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

Review · August 31, 2026

An Attenuated Delegation Profile for Automated Agents

draft-hamr-oauth-agent-delegation -00View on Datatracker →

The problem it's solving

A person asks an agent to rent a car. That agent hires a second agent at a different company. The second agent calls the rental service's API. The rental service is now two hops from the human and has a relationship with none of the parties in between.

The draft argues that service needs three things: proof the authority traces back to the person, proof nobody along the way gave themselves more than they were given, and the ability to check both without calling anyone.

The unique angle is the second one, checked strictly. Most attenuated-delegation drafts define a token format that supports narrowing. This one says the format is your business and the rules are the specification. It defines an Agent-Delegation HTTP header, three attenuation rules, and an ordered verification procedure, and it is explicit that a verifier which checks only the last link has defeated the whole thing.

The second thing it adds is "floors." A floor is a set of conditions about the human behind the chain, delivered to the verifier as a signed boolean rather than a raw value. Nobody downstream learns the account age, only that it cleared the bar.

How it fits in the working group

The draft quotes the OAuth charter's Complex Delegation work item and says it targets it directly. That quote is accurate, and cross-domain agent delegation is squarely in the WG's current program.

The rest of the fit is stranger. This draft cites no OAuth specification. Not RFC 6749, not Token Exchange, not RFC 9068, not RAR, not RFC 9700. Its four normative references are BCP 14, HTTP Message Signatures, Structured Fields, and HTTP Semantics. There are no OAuth roles in it. No client, no authorization server, no access token. It defines Delegator, Delegate, Relying Service, and Attestation Issuer from scratch.

So mechanically this is an HTTP header specification that happens to be about delegation. It would sit as naturally in HTTPAPI or WIMSE as in OAuth.

That matters most where RFC 8693's act claim is concerned. The known limitation of act is exactly the problem this draft opens with: it records prior actors but doesn't enforce anything across a chain. That's the obvious starting point for this conversation and the draft never mentions it.

The overlap the WG should actually look at is draft-asor-wimse-agent-delegation-chain, filed four days earlier. Same problem statement, same monotonic attenuation, same offline verification, same bounded depth. The difference is that asor picks RFC 9068 JWTs and RAR and defines a cryptographic commitment to the parent link. This draft picks nothing. The relationship section lists asor as a neighbor and declines to compare, which is a gap the WG will have to close itself since these two cannot both be adopted as-is.

The positioning against draft-klrc-aiagent-auth is better handled. The draft claims it fills two extension points klrc leaves open, and Appendix B records a genuine conflict between klrc's stable-identifier requirement and this draft's unlinkability requirement without pretending to have solved it.

What the mechanism actually does

The header is a Structured Field List of Byte Sequences, ordered root first, leaf last. The immediate sender covers it with an RFC 9421 signature tagged agent-delegation.

Every adjacent pair in the chain must satisfy three rules. Scope only shrinks. Floors only tighten. Expiry only moves earlier. Scope is a set of opaque strings compared octet-for-octet, and the draft forbids wildcards, prefixes, namespace containment, case folding, Unicode normalization, and whitespace trimming, with a clear explanation of why: a verifier that invents a matching rule accepts chains another verifier rejects, and then attenuation isn't checkable at all.

Floors are five named axes and no registry: subjectClass, tenureMin, credentialAgeMin, accountClass, partialPolicy. Durations are restricted to P<n>D or P<n>Y, with months banned because a month's length is ambiguous and would make tightening comparisons undefined. Unknown axis names are a hard rejection, not a silent skip.

Attestations must be signed over exact presented bytes, echo the verifier's nonce, carry an expiry, carry only a boolean, and be rejected outright if the serialization has a duplicate top-level key. A refusal to attest must itself be signed, with the same nonce and expiry rules, so it can't be forged or stripped. There's no revocation mechanism at all, and the draft says so plainly: short expiry is the only mitigation offered.

Then there's the part the draft doesn't tell you until Appendix A. Those five generic-sounding floor axes are mobile network operator attributes. tenureMin is how long the line has been active. credentialAgeMin is time since the SIM was last swapped. accountClass is postpaid versus prepaid. The author has a companion API proposal filed to CAMARA's backlog for the operator side of this.

That reframes the document. The abstract requirements in the body are a generalization of one specific deployment, and the axis set is what gives it away.

Significant interoperability concerns

The link format is undefined. Section 4 says each list member carries "one opaque, encoded delegation link" and that this document does not mandate an encoding. Section 11 then requires the verifier to parse each member into a scope, a floor, an expiry, and a signature. So the specification is meticulous about comparing scope strings octet-for-octet while leaving open what bytes a link is made of. Two conforming implementations cannot exchange a single chain. Everything downstream of that is unimplementable in the sense that matters, which is between two parties who didn't coordinate in advance.

Nothing binds a link to its parent. The terminology section defines a Link as including "the signature that binds them to their issuer and to their parent link." No normative section specifies that binding. Section 9 lists five required attestation properties and a parent reference is not among them. Section 7 establishes parent-child purely by position in the header list.

That's an attack. Take L(0) from a chain you legitimately hold. Take L(1) from a chain someone else presented to a different service. Put them in a header in an order that satisfies the three rules. Step 4 verifies each link's own signature, which passes because both links are genuinely signed. Step 6 checks attenuation between adjacent list members, which passes because you chose the order. The chain verifies. Splicing links from unrelated chains is the failure mode a delegation chain exists to prevent, and asor's byte-commitment to the parent is exactly the missing piece.

Nothing binds the chain to the presenter, either. No link carries a key or identifier for the Delegate it grants to. Section 10's agent identifier is about how a Delegate names itself to a Relying Service, is not required to appear in any link, and the draft explicitly declines to say how it's derived. The RFC 9421 signature proves the immediate sender signed the request, but nothing ties the sender's key to the leaf link. Capture the header bytes, sign your own request with your own key, and the chain still verifies. This is a bearer credential with signatures on it.

The nonce requirement makes the protocol impossible to satisfy. Step 7 of the verification procedure requires that every floor attestation in the chain echo the nonce the verifier issued for this exchange. Walk through the draft's own example. L(0) is created on the person's device when they ask for a car rental. The rental service's nonce doesn't exist yet, and won't until Agent A has already engaged Agent B. The root attestation cannot carry it.

The worked example never shows a nonce, which is how this survived. The implementation status section confirms multi-hop was never built. Either attestations are re-fetched at every hop against a fresh nonce, which means the offline-verification property in the introduction is gone, or step 7 applies only to the leaf, which needs saying. As written the document requires something that cannot happen.

Uniform rejection fights operability. Every failure must produce an indistinguishable outcome so an attacker gets no oracle. That means clock skew, an expired chain, a malformed header, and a genuinely unauthorized agent all return the same thing, and a well-behaved client has no way to learn it should refresh. OAuth went the other way deliberately with invalid_token and insufficient_scope. The threat here is also thin: to probe past signature verification an attacker needs valid signatures at every link, at which point they're not guessing.

Floors have no usable values outside one industry. Five axes, closed set, no registry, extension declared future work. If you're an enterprise SaaS deployment there is nothing in that table you can put in a floor. partialPolicy isn't even a condition about a subject, it's a knob for how an operator's API behaves when it holds incomplete data, which has no business in a delegation link.

Mechanisms that could be removed

Floor tightening down the chain needs justification it doesn't get. Attenuating scope at each hop makes obvious sense. A floor is a condition about the human's account, so it's unclear why a sub-agent would raise the required SIM age, and unclear what a verifier should do with the fact that it did. The worked example has Agent A raise credentialAgeMin from 90 to 180 days and never says why an agent would. Drop rule 2's tightening and keep floors as a root-level requirement that must be carried forward unchanged, and the model gets simpler without losing anything the draft demonstrates a need for.

Section 9.1's requirement that a signed refusal be verified "along a verification path independent of the path used to verify an accepted attestation" is an instruction about how to structure your code, not an interoperable behavior. Nobody on the wire can tell whether you complied.

partialPolicy should go entirely.

Appendix A is three pages on mobile operators, zero-knowledge mdoc proofs, and ICAO 9303 passive authentication. The trust-model discussion in Direction 3 is genuinely good, and it's a blog post rather than spec text. Cutting it would also stop the appendix from quietly answering the question of what the floor axes are for.

Writing quality

Better than most first drafts. Clean BCP 14 usage, correct structured-field vocabulary, requirements written as conditions a verifier can test. The motivation section has the best line in the document: "A delegation chain that is only checked at its endpoint is not a chain; it is a single trust decision with extra steps."

Two habits work against it.

The first is defensive hedging on almost every claim. "This document does not claim." "Nothing in this appendix is required." "Neither position is presented here as superior to the other." "Nothing about it has been approved, accepted, endorsed, reviewed, or adopted." The instinct behind it is good, and at this volume it reads like the author is arguing with reviewers who haven't shown up yet.

The second is restatement. Section 3 restates Section 1. Section 16 restates 9.2, 11, and 13. Appendix C restates Section 7 and then explains at length why negative test vectors are useful. This is 28 pages carrying three attenuation rules, one header field, five attestation properties, and a nine-step procedure. It could be twelve.

There's also a terminology problem in the middle of the model. Section 11 step 5 says "verify the root attestation, L(0)," but Section 7 defines L(0) as the root delegation link. Section 7 also requires validating L(0) "against the root Attestation Issuer's published floor," and a "published floor" appears nowhere else in normative text. Link, floor, and attestation are three different things in the terminology section and blur together by the verification procedure.

RFC 8941 was obsoleted by RFC 9651 in 2024. Minor, but the structured-field references should move.

What the draft gets right about itself

This deserves its own section because it's rare.

Section 15 lists what the proof-of-concept does not implement, and the list includes the multi-hop delegation chain, which is the entire subject of the document. Appendix C states that the test vectors "have not been executed against any implementation by anyone, including the author." Section 9.2 exists solely to say that echoing a nonce is not replay protection and implementers must not assume it is. Section 16 says cost-based floors raise the price of a reset and establish nothing about uniqueness. Section 13 says the Attestation Issuer's query log is "a real and unsolved limitation of this profile, not an oversight to be corrected by better protocol design."

Most first drafts oversell. This one marks its own limits more carefully than some adopted work does, and the V6 test vector, built specifically to catch a verifier that only checks the leaf, is the kind of thing a careful spec author writes.

About the author

Amr Hassan, Independent, San Francisco. Sole author.

Zero published RFCs. One active Internet-Draft, which is this one. No expired drafts. New IETF participant with no prior document history. The companion CAMARA API proposal is also filed under the author's name.

Where the working group should push

The core claim is correct and worth keeping in the conversation whatever happens to this draft. Whole-chain verification is a real requirement, act doesn't provide it, and a chain checked only at the leaf is not a chain. The scope-comparison section is the strongest part of the document and the reasoning in it generalizes.

Four questions for the author:

  • How does a link commit to its parent, and how does a link commit to the Delegate it grants to? Without both, the chain can be spliced and the credential is bearer.
  • How can the root attestation carry the final verifier's nonce, given the root is created before any of the downstream parties exist?
  • What encoding does a link use? The three rules can't produce interoperable implementations while the thing they operate on is undefined.
  • What are the floor axes for a deployment that isn't a mobile operator? A closed set with no registry means the answer today is nothing.

There's a bigger call for the chairs. This draft and draft-asor-wimse-agent-delegation-chain are solving the same problem, filed four days apart, and neither cites a mechanism from the other's design. The WG needs one attenuated-delegation conversation rather than parallel drafts converging by accident, and given the OAuth-WIMSE coordination already in the charter, that conversation should probably happen across both groups at once.