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

Update review · November 10, 2026

KYAPay Token Update

draft-skyfire-oauth-kyapay-token -03View on Datatracker →

Status: Individual draft (not adopted), Standards Track. Revision -03 published 2 November 2026, one month after -02. 46 pages, up from 36. Same two authors.

This is a big revision. It reworks the identifiers, adds a pseudonymous human ID, defines how a target is named, and tightens validation. It fixes about half of the interoperability concerns in the previous review. None of the structural feedback was taken: PAY is still in the document, the 16 short claim names are still registered globally, and there is still no mention of Web Bot Auth.

What got fixed

  • verified: true now has detail behind it. hid gained ivm, amr, and aml sub-claims, and apd gained ivm and aml. These come from the sibling Skyfire drafts, which is what the previous review asked for. They are banned as top-level claims in a KYAPay token.
  • A target can validate the audience without registering at the issuer. tdm is now required when the target has a domain name. A recipient that hasn't onboarded checks tdm instead of aud. The format is defined: lower-case host plus optional path, with matching rules borrowed from cookie paths.
  • hid is plainly REQUIRED. The self-contradicting qualifier is gone.
  • hid.email is OPTIONAL. So are all the other directly identifying sub-claims.
  • creation_ip is OPTIONAL and has a defined purpose. It's a diagnostic aid. A recipient "MUST NOT treat it as sufficient on its own."
  • sub is now actually pairwise. It's a different value at each target service, which matches what "pairwise" normally means.
  • Issuer trust is checked before key retrieval. A recipient MUST NOT fetch a JWK Set for an issuer outside its configured set. In -02 the signature check came first, which meant fetching keys from whatever iss the token named.
  • ES256 is mandatory to implement. -02 only recommended it. Symmetric algorithms and none are rejected outright.
  • The payment examples no longer look like real card numbers. The -02 example token also had "verified": false in hid. That's corrected.
  • RFC 8725 is now a MUST.

What didn't

  • No trust model. The set of trusted issuers "is configured out of band." That's now explicit, but it's the same answer as before.
  • PAY is still here. The offer to split it out "if desired by a working group" is unchanged.
  • Payment semantics are untouched. val is still "settlement network's units." pay_per_mb and subscription are still undefined. Nothing covers capture, refunds, or mnr running out.
  • All 16 claims are still top-level registrations, including env, cur, amt, val, and ori.
  • /.well-known/jwks.json is still claimed for KYAPay signing keys.
  • source_ips still has no processing rules.
  • env, itg, ori, tpr, tps, and mnr are all still there with the same one-line definitions.
  • The verifier/verified/verification_id triplet stayed. The new method claims were added next to it.
  • No Web Bot Auth, no act, no verified_claims.
  • The editorial list is untouched. The intro still says agents "bypass common blocking mechanisms." The privacy section still says "If I authorize an agent." PAY validation still says "the price that you configured." Validation still requires jti to be a UUID and the claim definition still doesn't. Both typos survive. Mike Jones is still the change controller on every registration.

What's new

Pseudonymous identifiers. hid.id is a new required opaque ID for the human. It's specific to one agent platform and one target service. aid.id is a new required ID for the agent, specific to one target service. sub now identifies the combination of human, platform, and agent at one target service. A recipient can use sub to rate-limit one agent without affecting the same user's other agents.

Data minimization. A KYA token now carries only hid.id by default. Email, names, and phone are included only when the target requires them and the user has authorized it. This is a real change in posture. The -02 token put the user's email in every request.

Identifying the Target (new Section 3.1.1). aud and tsi are issuer-assigned and only meaningful to onboarded targets. tdm is public. tsi is now REQUIRED. A CDN or bot manager protecting many sites checks that tdm matches the registrable domain, using the Public Suffix List.

Token Size. Issuers SHOULD keep claims under 4096 octets. One KYA-PAY token is preferred over a KYA and a PAY token together.

typ handling. A recipient decides which of the three types it accepts per request. A valid type it doesn't want is "set aside" instead of failing validation.

Identity Binding (new Section 5.6). A recipient correlating a PAY token with a KYA token must check that iss and sub both match.

PCI scope argument. Card credentials are now defined as EMV Payment Tokens, with a citation to PCI SSC FAQ 1326 saying they are out of PCI DSS scope. The ban on PANs now extends to track data, PINs, and all three token types.

New problems

apd is both REQUIRED and OPTIONAL. Section 3.4.2 says "The apd claim is REQUIRED in KYA and KYAPay tokens." The claim list one page earlier still says "OPTIONAL - Agent Platform identity claims." The sub definition says tokens carry apd "where it is identified."

Issuers now mint tokens for targets they've never verified. Where the issuer can't confirm the target, aud and tsi are "its record of the Target that the initiator asserted," and tdm is whatever domain the initiator supplied. That's fine for audience restriction. But tsi is REQUIRED and "assigned by the Token Issuer," and the draft doesn't say how an issuer assigns a stable service ID to a site it has no relationship with. All three pairwise identifiers are keyed on tsi, so this matters.

The aud definition is hard to follow. It "identifies the combination of identities the issuer holds for the Target: the human principal, the agent platform and the agent." For a website, that's three roles mapped onto one thing. The examples still label it "Target Agent Account ID."

The pairwise identifiers don't prevent linking. The draft admits this. apd is the same at every target. aid.name is the same at every target and must be unique per platform. A personal agent with one user links that user everywhere through its name. source_ips and creation_ip do the same. The privacy claim holds only for agents shared by many users.

Replay detection got weaker. -02 said jti "can be used to detect replay attempts." -03 says a repeated jti "is not by itself evidence of replay," because a token can be presented more than once. For a bearer KYA token, that leaves exp and audience as the only defenses. For a PAY token carrying a single-use cryptogram, the draft doesn't say whether the token itself is single-use.

verified is a boolean that can only be true. A recipient that sees false treats it as absent but doesn't reject the token. The claim now carries no information that ivm doesn't carry better.

Normative references to three individual drafts. The ivm, amr, and aml drafts moved from informative to normative. Meanwhile the Using KYAPay Tokens draft is still informative, even though cnf processing, the minimum token size, and the tdm validation choice are all delegated to it. The draft can't be implemented without it.

Domain matching depends on the Public Suffix List. Two recipients with different copies of the list will disagree about which hosts a token covers. A domain-only check also means a token minted for shop.example.com passes at every other host under example.com. The draft says so and leaves the host and path check to the target.

Background of the Authors

No change. Ankit Agarwal and Michael B. Jones are still the only authors. See the previous review for their backgrounds.

Suggested Feedback

  1. Acknowledge the progress on hid. Pseudonymous by default with disclosure on request is a much better design than email in every token. The ivm/amr/aml integration and the issuer-before-keys ordering are also direct fixes.

  2. Fix the apd contradiction. Pick REQUIRED or OPTIONAL and make Section 3.4, Section 3.4.2, and the sub definition agree.

  3. Say how tsi is assigned for a target that hasn't onboarded. If it's derived from tdm, say so. If it isn't stable, then sub, hid.id, and aid.id aren't stable either.

  4. Make aid.name pairwise or drop the uniqueness requirement. As written, it undoes the unlinkability that hid.id was added to provide.

  5. State whether a PAY token is single-use. Then say what a recipient does with a repeated jti on one.

  6. Drop verified. It can only be true, and ivm says more.

  7. Repeat the structural asks. Split out PAY. Nest the claims under one or two registered names. Use RFC 8414 metadata instead of claiming /.well-known/jwks.json. Talk to the Web Bot Auth WG. None of these were addressed, and the document grew by ten pages around them.

  8. Move Using KYAPay Tokens to a normative reference, or pull the cnf and size requirements into this draft.

For related reading, see the companion Using KYAPay Tokens draft, which now defines most of the recipient behavior this draft points to.