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

Update review · November 20, 2026

An Attenuated Delegation Profile for Automated Agents -04 Update

draft-hamr-oauth-agent-delegation -04View on Datatracker →

Status: Individual draft (not adopted), Standards Track. Revision -04 published 15 November 2026, about three weeks after -03. 38 pages, up from 37. Same sole author.

This revision changes the definition of x and rewrites the classifier benchmark. The delegation chain is untouched. Sections 4 through 7 and 10 through 13 are identical to -03. Of the nine interoperability concerns in the -03 review, one was partly addressed and one got worse.

What got fixed

  • Granting access is now x. The -03 review pointed out that resetting another user's password was w. Section 9.1 now says a call that grants or widens access is x: a role binding, an ACL, a share, an invitation, a deploy key. Setting another person's password is x. A call that sets access without saying which direction it moves is x.
  • The Jev figures are gone. The optional language model pass couldn't be reproduced and used the old truth labels. It was removed from Section 16.

What didn't

  • The link format is still undefined. Section 4 still says the draft does not mandate an encoding. Two implementations still can't exchange a chain.
  • Nothing binds a link to its parent or to the presenter. Parent and child are still established by position in the list. A captured chain still works for anyone.
  • There is still no key model for link signers.
  • The attestation nonce still can't be delivered.
  • The interim deployment still needs no standard. The agent's operator still signs the menu, issues the link, and runs the verifier.
  • Budgets still leak. A new root link still gets a fresh count.
  • "Published floor" is still undefined.
  • None of the removal suggestions were taken. The person floors and action classes are still one document. Section 12, Section 9.6, the layer taxonomy, and Appendix A are all still here.
  • The accuracy figure is still in the Introduction. It is a different number now, and it gained a second citation.
  • Token Exchange and act are still never mentioned, five revisions in.
  • Appendix B still points to the expired -01 for the full test vectors.
  • Nothing implements the protocol. The justabit suites are still dated 22 October. No code runs the test vectors.

What changed

The definition of x is broader. In -03, x was a list of verbs. In -04 the test is whether a later call can put things back the way they were, "including anything anyone could have done in between." The access rules above follow from that. A policy that doesn't control access, such as autoscaling or retention, stays w.

QUERY defaults to r. Under -03 it fell under "any other method" and defaulted to x. RFC 10008 is a new normative reference.

The benchmark was replaced. rwxmap's rules were rewritten. A new classifier, rwxlite, is a fine-tuned all-MiniLM-L6-v2 sentence model that rwxmap can hand marked calls to. The test set is now 1588 calls from Grafana, NocoDB, Radarr, Cloudsmith, and Invoice Ninja. It was 4279 calls from Cloudflare, PagerDuty, and Sentry.

No protocol mechanisms were added. None were removed.

New problems

The class definition changed again. That is four revisions in a row. The -03 review asked why the working group should believe r/w/x is stable. A PUT on a role binding was w three weeks ago and is x today. Any menu signed under -03 is wrong under -04, and nothing in a menu says which definition it was written against.

The new test has no clear edge. "Anything anyone could have done in between" applies to most writes. Someone can read an edited document, act on a toggled feature flag, or route traffic through a changed rule before it is set back. The draft says granting access is x for this reason and autoscaling is w despite it. The line between them is a judgment call that two menu signers will make differently.

Narrowing access is w, and the reference tool disagrees. "Revokes" was in the -03 list for x. It is gone in -04, and Section 9.1 says a call that only narrows access is w. Section 9.5 and Section 16 still describe rwxmap raising a GET to x when its name is "revoke." Removing the last admin from an account also only narrows access, and no later call by anyone can undo it.

The method default is now wrong by the draft's own definition. PUT and PATCH default to w. Role bindings, ACLs, and deploy keys are commonly set with PUT. Section 17 still only warns that the default can be too loose for GET. A verifier using classSource of method admits an access grant under a link limited to w.

The headline leak rate went up. The released default classes 2.2% of calls looser than the truth, against 0.8% in -03. The draft says the two can't be compared because x and the labels both changed. Within -04's own table, the rules alone leak 2.7% and the classifier alone leaks 0.8%. Combining them gives the best accuracy and not the fewest leaks. A leak is a call a floor should have refused.

The benchmark limits are disclosed and they are serious. The truth labels came from two copies of the same language model, with the author ruling on 24 disputes. The draft says this shows consistency and not truth. The "unseen" check compares provider names, and some training descriptions mention these providers. The int8 and fp32 builds of the classifier "do not always agree," so the menu a harness signs depends on which build it ran.

More product detail in a Standards Track document. Section 16 now gives model weight counts, file sizes, training set sizes, npm versions, and a confidence threshold. The -03 review called the classifier figures blog material. There is more of it now.

Background of the Author

No new authors. Amr Hassan, Independent, is still the sole author.

The Datatracker profile loaded for this review. It shows zero published RFCs, one active Internet-Draft (this one), and no expired drafts. That is a new IETF author. The draft has had five revisions since 31 August.

Suggested Feedback

  1. Freeze the class definition or version it. Put a definition version in the menu, so a verifier can tell which reading of x the signer used.
  2. Fix the method default for access-granting calls, or say it can't be fixed. Section 17 needs a PUT and PATCH warning to match the GET one.
  3. Reconcile "narrowing is w" with revoke. Say what class a token revocation and a last-admin removal get.
  4. Move the benchmark out. Section 16 can cite the tools. The figures don't belong in Section 1 or Section 9.5.
  5. Repeat the open items from -03. What encoding does a link use? How does a link commit to its parent and to the Delegate presenting it? How does an attestation get the verifier's nonce?
  6. Split the document. The action class label and signed menu are the useful part, and they are the only part that moved in this revision. The delegation chain has not changed since -03 and its core questions date to -00.

For related reading, see draft-asor-wimse-agent-delegation-chain, which defines the token format and parent-link commitment this draft still lacks.