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

Update review · November 16, 2026

KYAPay Token -04 Update

draft-skyfire-oauth-kyapay-token -04View on Datatracker →

Status: Individual draft (not adopted), Standards Track. Revision -04 published 16 November 2026, two weeks after -03. 49 pages, up from 46. Same two authors. It is on the IETF 127 OAuth agenda.

This is a small revision. It fixes one of the eight items in the -03 review, adds a record-keeping section for issuers, and adds an auth_time sub-claim. The token format is otherwise the same. None of the structural feedback from the original review was taken.

What got fixed

  • apd is REQUIRED everywhere. The claim list in Section 3.4 said OPTIONAL while Section 3.4.2 said REQUIRED. Both now say REQUIRED. The sub definition dropped "where it is identified" (Section 3.1).
  • aid is stated to be the agent, not the human. Section 3.4.3 now says claims about the human principal belong in hid. This is a clarification, not a behavior change.

What didn't

  • tsi for a target that hasn't onboarded. Still "assigned by the Token Issuer" with no rule for a site the issuer has no relationship with (Section 3.1.1).
  • aid.name still links. It is still required, unique per platform, and the same at every Target. Section 7 still admits it links a single-user agent everywhere.
  • PAY single-use is still unstated. A repeated jti is still "not by itself evidence of replay" (Section 3.1).
  • verified is still a boolean that can only be true (Section 3.4.1).
  • Using KYAPay Tokens is still an informative reference (Section 9.2), and cnf processing is still delegated to it.
  • PAY is still here. Same offer to split it "if desired by a working group."
  • All 16 claims are still top-level registrations. /.well-known/jwks.json is still claimed (Section 8.3).
  • No trust model. Trusted issuers are still "configured out of band" (Section 4.1).
  • No Web Bot Auth, no act, no verified_claims.
  • Payment semantics and source_ips processing are untouched. pay_per_mb and subscription are still undefined.
  • The editorial list is untouched. "Bypass common blocking mechanisms," "If I authorize an agent," "the price that you configured," both typos, and Mike Jones as change controller all survive. The examples still label aud "Target Agent Account ID."

What's new

Issuer Records (new Section 5). An issuer MUST keep three kinds of records. Verification records cover the methods, times, and evidence behind each verification. Authorization records cover which sub-claims the principal agreed to disclose, to which Target, and how. Issuance records cover each token's iss, jti, iat, exp, sub, aud, tsi, and the three pairwise IDs. The stated goal is that iss plus jti resolves a token to everything behind it. Retention periods and disclosure to courts or regulators are left to local law.

hid.auth_time (Section 3.4.1). Optional. It is the OpenID Connect claim, nested inside hid. When a token is issued from an earlier authentication, it carries the earlier time. A recipient that wants a recent confirmation compares auth_time, not iat. Like amr, it is banned at the top level and in apd.

Pairwise per Target, not per service. sub, hid.id, and aid.id must now differ "at each other Target." -03 said each other Target service. Whether two services of one Target share a value "is left to the issuer."

Token retention. -03 banned writing tokens to any store not needed to process the transaction. -04 adds an exception for audit, dispute resolution, and legal retention (Section 6.2).

New problems

Section 5 is unverifiable from the outside. It puts a MUST on an issuer's internal bookkeeping. The draft says it "does not define an interface for querying" the records. A recipient can't test compliance and can't use the records. This is policy for an issuer network, placed in a token format spec.

The issuance log is a map of every user's activity. The issuer must record every token with its Target and all three pairwise IDs. The draft says to protect the records. It sets no retention limit.

The pairwise change weakens a privacy promise without defining the boundary. "Target" is never defined as one thing. Section 2.1.2 defines a Target Agent, Platform, Principal, and Identity. A recipient also can't tell whether an issuer shares IDs across a Target's services.

The retention exception covers PAY tokens. A logged KYA-PAY token contains the card network token and cryptogram. "Legal retention obligations" is a wide exception to put next to a settlement credential.

auth_time has two meanings. For a token "issued for a specific action," it is the time the human confirmed that action. Otherwise it is the last login. The definition gives the recipient no way to tell which one it got.

Background of the Authors

No change. Ankit Agarwal and Michael B. Jones are still the only authors. Ludovic Dépinoy and Rob Zagarella were added to the acknowledgments.

Suggested Feedback

  1. Ask what Section 5 is for. Either define how a recipient or auditor queries the records, or move it to an operational document.
  2. Define "Target." The pairwise identifiers now depend on a term the draft doesn't define.
  3. Exclude settlement credentials from the retention exception.
  4. Repeat the open items from -03. tsi assignment, aid.name, PAY single-use, verified, the normative reference, and the editorial list.
  5. Ask for an answer on the structural items at IETF 127. Split out PAY. Nest the claims. Use RFC 8414 metadata. Talk to the Web Bot Auth WG. Two revisions have now gone by without a response to any of them.

For related reading, see the companion Using KYAPay Tokens draft, which was revised the day before.