Update review · November 10, 2026
Using KYAPay Tokens Update
Status: Individual draft (not adopted), Standards Track. Revision -02 published 2 November 2026. 45 pages, up from 32. Same three authors.
This revision fixes most of the interoperability concerns in the previous review. It adds an error response, starts an HTTP Message Signatures profile, and stops keying account access on email. None of the suggested cuts were made. The document grew by 13 pages and still never mentions Web Bot Auth.
What got fixed
- There is now an error response. A recipient returns 403, or 402 when payment is missing. It can add a new
Accept-KYAPay-Tokenresponse header listing the claims it needs. The agent is no longer blocked with no explanation. - Account linking is no longer keyed on email. A CIAM system MUST establish that the token's principal is linked to an existing account before admitting an agent. The link is the stored
issplushid.id. "A validated token alone MUST NOT be treated as sufficient." - The header is a real Structured Field.
KYAPay-Tokenis an RFC 9651 List of Tokens. The custom ABNF is gone. A field that fails to parse rejects the whole request. - "Admit by default" is gone. The intro now says "identification is distinct from admission." Admission depends on the site's policy for the requested action.
- Cross-site correlation is no longer a feature. Linking uses
hid.id, which is different at every target service. Email appears only when the target asked for it and the user agreed. - The
envandcnfchecks are separate steps.
What's partly fixed
Proof of possession has the start of a profile. New Section 4.4 says the signature MUST cover @method, @target-uri, the KYAPay-Token header, and Content-Digest when there is content. It MUST carry created and expires. The first hop verifies it before any URL rewriting. That is real progress.
The rest is deferred. The key source, keyid, tag, and error responses "are expected to be specified in a future revision." There is no nonce, and the draft accepts that an identical signed request can be resent until the signature expires. Two implementations still can't interoperate on this.
Email matching is still a permitted link. An account is also linked when "a verified identifier in hid matches one the account records as verified." That is the same email match as before, now resting on the issuer having verified control. For ATO protectors the link check is only a SHOULD.
What didn't change
- The normative reference is still stale. It cites
draft-skyfire-oauth-kyapay-token-02. The draft depends onhid.id, requiredtdm, the 4096-octet claim limit, and the "Identifying the Target" section. All of those are in -03. - Informative sections still contain MUSTs. Section 4 still opens with "This section is informative." Section 4.4 sits inside it and is full of MUSTs that the processing model depends on.
- "Hard limits enforced by the token" is still there. A JWT doesn't enforce a payment limit.
- Nothing was removed. The Design Principles, the role-by-role guidance, Feedback to the Token Issuer, and Token Creation are all still present. The Bot Managers section grew into three subsections.
- No Web Bot Auth. The draft now profiles RFC 9421 for bots at the edge and still doesn't mention the working group doing the same thing.
- RFC 9470 was considered and rejected. The draft cites it for a privacy point, then says recipients SHOULD NOT use 401 or
WWW-Authenticate, because a KYAPay token "is not an Authorization credential." - The
envcheck stayed. - The small issues stayed. The policy example still uses undefined values like
document+biometric. Steps 3 and 4 of the processing model still both verify the signature. NIST SP 800-63A is still revision 3. "Verifier" is still not defined.
What's new
Accept-KYAPay-Token (Section 5.6). A Dictionary Structured Field. Each member is a recipient-chosen label with an inner list of claim names, like checkout=("hid.email");purpose=checkout. Three parameters are suggested: purpose, optional, and issuers.
Invalid tokens are now declined. In -01 a failed validation fell back to normal bot detection. Now the recipient MUST decline: the bot manager blocks, the fraud manager declines the transaction, the CIAM system refuses the account action.
Attaching Tokens to Requests (Section 5.5). The token goes to the target and to hosts acting for the target. Other hosts SHOULD NOT get one. Vendor hosts with no configured claims get a "base claims" token with only hid.id, assurance claims, apd, and aid.
Token size. A recipient MUST accept a header value of at least 6144 octets and SHOULD answer 431 when it rejects one for size.
Tokens are reusable. A recipient MUST NOT treat a repeated token as replay. jti is for correlation and audit.
Audience checking by tdm. A CDN or bot manager that hasn't onboarded with the issuer checks that tdm names the domain it protects.
Token Exposure in the Agent Runtime (Section 8.6). This is the best new section. It covers tokens leaking through model context, transcripts, tracing, and prompt injection. The component that attaches a token MUST decide without relying on model output.
HTTP details. CORS preflight requests are exempt from token-absent policy. Responses that depend on the token need Vary: KYAPay-Token and must not be stored in shared caches.
New problems
Accept-KYAPay-Token can't express assurance. It lists claim names. A site that needs document-plus-biometric verification can only ask for hid. The agent gets a token, retries, and gets another 403. Step-up was the main reason to have this header.
Its parameters aren't interoperable. "Recipients define the parameters and values they need." purpose has no defined values, so an agent can't show it to a user in any reliable way. issuers is a space-separated string inside a Structured Field that already has lists. The problem types in the examples are "for discussion only."
SHOULD and MUST disagree on signing. Section 5.1 says a token with a cnf key "SHOULD be accompanied by" a signature. Step 7 of the processing model says the recipient MUST reject the request if the signature is absent.
Sending a token can be worse than sending none. A failed token is declined. A missing token gets normal bot detection. The draft doesn't say which path applies when the issuer simply isn't on the recipient's list. An agent with a valid token from an issuer the site hasn't configured may get blocked where an anonymous bot gets through.
The agent has to know the target's vendors. Section 5.5 sends tokens to third-party hosts "acting on its behalf." The agent learns which hosts "by configuration or from an arrangement with the Target." That contradicts the no-per-site-setup principle in Section 3.2. It also sends a bearer token audienced to the target to more parties.
Page scripts can't read request headers. The fraud manager section suggests "script running on the Target's page captures the KYAPay-Token value." A script has no access to the headers of the request that loaded its page.
"Human presence signal" is now more wrong. The definition of a human-via-agent request now says the principal "need not be present." The signal is still named for presence.
The bot manager section reads like adoption marketing. A bot manager that detects automation without a token "should issue a custom challenge response" that "gently encourages developers to adopt and implement the KYA standard."
Background of the Authors
No change. Ankit Agarwal, Michael B. Jones, and Srinivasa Thumma are still the authors. See the previous review for their backgrounds.
Suggested Feedback
Acknowledge the fixes. The error response, the
hid.idaccount binding, the Structured Field definition, and Section 8.6 are all direct improvements.Finish the RFC 9421 profile or point to one. Key source,
keyid,tag, and errors are the parts implementers need. The Web Bot Auth WG has already made these choices for the same use case.Let
Accept-KYAPay-Tokenexpress assurance requirements. Define how a recipient asks for a minimumivm,amr, oramlvalue. Define the parameters, or register them.Say what happens with an untrusted issuer. Treating it as "no token" avoids punishing agents for identifying themselves.
Update the normative reference to kyapay-token-03.
Move Section 4.4 out of the informative section and make Section 5.1 agree with step 7.
Drop the email match as a sufficient account link, or require the recipient's own verification on top of it.
Repeat the cuts. Sections 3, 6.3 to 6.10, and 7 are still product guidance. The header, the response field, the signature profile, and the processing model would fit in about ten pages.
For related reading, see the KYAPay Token draft, whose -03 revision defines the hid.id and tdm behavior this draft now depends on.