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

Review · October 26, 2026

Using KYAPay Tokens Summary

draft-skyfire-oauth-using-kyapay-tokens -01View on Datatracker →

This draft goes with the KYAPay Token format. It defines a KYAPay-Token HTTP header for carrying the token. It defines a "recipient" role for the party that validates it. Most of the document then describes how bot managers, fraud managers, account-takeover (ATO) protectors, and CIAM systems should use the token's claims to make decisions.

The Problem

Bot managers identify automated traffic by IP ranges, ASNs, and User-Agent strings. Those signals show where traffic comes from. They don't show which human authorized it. An agent buying something for a real customer looks the same as a credential-stuffing bot.

The draft reframes the question. Bot managers today ask "is this a human?" This draft has them ask "did a verified human authorize this agent?" The agent sends a signed KYAPay token in the request. Every security layer reads that token instead of guessing.

The new co-author is the most important change. Srinivasa Thumma is from Akamai, which is a real bot management vendor. The token format alone was a Skyfire spec. This draft shows interest from the people who would actually consume the token.

How It Fits In

This is not an OAuth protocol. It defines no OAuth endpoints, grant types, or parameters. Its only normative OAuth dependencies are Token Exchange (RFC 8693) and RFC 7800 for cnf. It mentions DPoP, mTLS, and Client ID Metadata Document only in passing.

The main mechanism is an HTTP header that bot managers at the edge read. That is the Web Bot Auth WG's scope. The draft uses HTTP Message Signatures (RFC 9421) for proof of possession, which Web Bot Auth already profiles. Like the token draft, it never mentions Web Bot Auth.

The OAuth WG does come up in the ATO and CIAM sections. Those say a recipient can exchange the KYAPay token for an access token through KYAPay Token Exchange or RFC 8693. That is a reasonable place for OAuth work. It lives in a third draft.

What It Defines

The header. KYAPay-Token holds one or more JWTs as a comma-separated list. Recipients use the typ header to tell KYA, PAY, and KYA-PAY tokens apart. The header must only be sent over TLS. It must not be cached, logged, or forwarded.

A common processing model. A recipient:

  1. Checks iss against a trusted-issuer list configured out of band, before fetching anything.
  2. Validates the token per the token spec.
  3. Fetches keys from the issuer's /.well-known/jwks.json.
  4. Checks aud (and tdm/tsi if used).
  5. Checks env.
  6. If cnf is present, requires an HTTP Message Signature from that key.
  7. Falls back to normal bot detection if any step fails.

Step 1 is a good addition. It prevents a request with an attacker-controlled iss from triggering an outbound fetch.

Per-entity assurance. Recipients should treat the human (hid), the agent platform (apd), and the agent instance (aid) as separate entities. Each gets its own policy, with no single yes/no answer. A site can require document-plus-biometric verification on hid for /checkout and accept less for browsing. The same three tiers set how wide a block is. You can throttle one abusive user without blocking the whole agent platform.

Role guidance. There's a subsection each for bot managers, fraud managers, ATO protectors, and CIAM systems. Most of it is SHOULD-level advice about which claims to feed into risk scoring.

Proof of possession as "planned evolution." Tokens are bearer tokens today. When cnf is present, it must carry the full jwk and the agent signs each request with RFC 9421. The draft explains why it picked this over a DPoP-style detached proof. The recipient validates one token per session, caches the key, and then only checks a signature on each request.

Significant Interoperability Concerns

Proof of possession isn't profiled. "MUST verify that the request is signed by that key" isn't enough to build to. RFC 9421 needs a profile. Which components are covered? Which signature label? How is keyid tied to cnf.jwk? What's the freshness window for created, and is a nonce required? Two implementations will get different answers. Web Bot Auth has already answered these questions for the same use case.

There's no error or challenge response. The draft never says what a recipient sends back when a token is missing, invalid, from an untrusted issuer, or below the required assurance. The agent gets blocked and has no idea why. Step-up is called "an open item." Step-up Authentication Challenge (RFC 9470) already defines a WWW-Authenticate error for exactly this.

Account creation and linking keyed on email. The CIAM section says a recipient MAY create an account from the hid.email claim. The ATO section has a token exchange pull "principal claims such as the email address from hid" to issue an access token for an existing account. That is the classic email-based account takeover. Say an issuer's verified: true was a weak email check. An agent holding that issuer's token could then get into someone else's existing account at the target. The draft needs to say plainly that matching existing accounts by a third-party email claim is dangerous.

The normative references are stale. It cites draft-skyfire-oauth-kyapay-token-01 and its "Section 4" validation. The JWKS discovery and cnf processing it relies on were added in -02.

Informative sections contain MUSTs. Section 4 says it "is informative." Section 4.3 then says the token "MUST carry cnf.jwk" and recipients "MUST NOT be required to obtain the key by any means other than reading cnf.jwk." Readers can't tell whether this is a requirement or a plan.

"Admit by default" contradicts "identification is not admission." The design principles say the goal "can only be met if security intermediaries admit legitimate human-via-agent requests by default." Section 6 says a valid token is "not an automatic grant of access." A bot management vendor reading this won't know which one is the requirement.

Header registration doesn't match its syntax. The IANA entry declares a Structured Field List. The draft then defines its own ABNF and doesn't reference RFC 9651. Choose one.

Correlating users across sites is presented as a feature. Section 6.11 points out that hid.email is stable across touchpoints. That lets targets link an agent's activity to a later in-store purchase. The token spec makes hid.email required. So every target, and every intermediary in front of it, gets a global identifier for the user on every request. The privacy section's "SHOULD prefer the least revealing set of claims" can't be met when the claim is required.

"Hard limits enforced by the token." The fraud manager section says the PAY token "pins amount, currency, and target," so fraud systems can treat those as enforced limits. A JWT doesn't enforce anything. The settlement network does, or doesn't.

Mechanisms That Could Be Removed

  • Most of Section 6.3 to 6.9. These describe how to build a good bot manager or fraud engine. "An operator SHOULD configure them to reach consistent verdicts" is product guidance, not something that affects interoperability. Reduce it to the processing model plus a short informative section.
  • Section 6.10, Feedback to the Token Issuer. The mechanism, format, and trust model are all left undefined. Take it out until there's something to specify.
  • Section 7, Token Creation. By its own description it is non-normative. Advice like "a REST endpoint for programmatic agents, several narrow MCP tools for LLM agents" belongs in implementation docs.
  • Section 3, Design Principles. Cut it down to two sentences. It's the source of the "admit by default" contradiction.
  • The env check. It's the same issue as in the token draft. Use separate issuers or keys for sandbox.

Writing Quality

The normative parts are clear: the header definition, the processing model, and the security considerations. The security section is honest. It names compromised agent hosts as an "irreducible" risk and calls issuer trust "an open problem."

The rest reads like a position paper. Design principles talk about "probabilistic discovery" of new merchants and "zero merchant tech lift." Section 4 calls identity "verifiable nouns" and signatures "a verb." The intro is better than in -00. It now presents agents as the next step after verified bot traffic, which is the right pitch for bot management vendors.

Smaller issues:

  • The policy example in 6.2 uses values like document+biometric and "registered by its platform" that aren't defined anywhere.
  • Steps 3 and 4 of the processing model overlap. Signature verification is already part of the token spec's validation.
  • Step 6 combines the env check and the cnf signature check into one step.
  • The document history says -01 "clarified the meaning of Verifier." The terminology section doesn't define Verifier.
  • NIST SP 800-63A is cited as revision 3. Revision 4 is current.
  • "Human presence signal" is the wrong name. The draft admits in 8.8 that the token shows the human authorized the agent at issuance, not that the human is present.

At 32 pages, the header and processing model would fit in about five. The rest is positioning and role-by-role advice.

Authors

Datatracker wasn't reachable from this session. This section is based on the drafts in the repo and on published history.

  • Ankit Agarwal (Skyfire Systems): New to the IETF. Appears only on the Skyfire family of drafts, all of which are active right now.
  • Michael B. Jones (Self-Issued Consulting): Author of dozens of RFCs, including JWT, JWS, JWK, Token Exchange, and DPoP. Co-author on every Skyfire draft and has many other active drafts.
  • Srinivasa Thumma (Akamai): Appears to be new to the IETF. Also co-author of the Skyfire amr values and ID verification drafts.

Bottom Line

This is the draft the KYAPay token format was missing. It gives the token a header, a recipient role, and a processing model. With Akamai involved, the bot management use case is real.

Few of its 32 pages define interoperable behavior. The parts that do are underspecified. Proof of possession needs an RFC 9421 profile. Recipients need a way to tell the agent why it was rejected. Account creation or login based on a third-party email claim needs a strong warning, not a MAY.

Feedback should push for four things. Take the header and signature profile to Web Bot Auth, or align with what that WG already defined. Reuse RFC 9470 for step-up and errors. Cut Sections 3, 6.3 to 6.10, and 7 down to informative notes. Keep the OAuth WG involvement to the token exchange piece.

For related reading, see the KYAPay Token draft and the Web Bot Auth WG.