| Internet-Draft | Accountable Principal References | November 2026 |
| Molas | Expires 22 May 2027 | [Page] |
When a software agent makes a request on behalf of a person or an organization, the party that receives it often needs to know that someone answers for the request, without needing to know who. Several specifications in progress carry some reference to the principal behind an agent, and they make different choices about who learns the principal's identity, when the reference can be linked across services, who can resolve it, and how fast a withdrawn principal stops being accepted. This document states six properties of such a principal reference in testable terms, maps each to existing specifications that already provide part of it (OAuth 2.0 Token Exchange, the JWT profile for OAuth 2.0 access tokens, OAuth identity chaining, Token Status List and Selective Disclosure for JWTs), and gives a first-pass evaluation of five current drafts against them. It defines no token format, protocol or claim, and it does not rank the drafts.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 22 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Agent identity and principal identity are separate questions. OAuth 2.0 Token Exchange already separates the acting party ("act", Section 4.1 of [RFC8693]) from the subject of a token, and Section 10.3 of [AIMS] conveys the agent in "client_id" and the user or system it acts for in "sub".¶
Many relying parties do not need the principal's identity to decide. A bot manager, an API or a merchant may only need to know that an established principal stands behind the request, at what assurance, and that it is the same principal as earlier in the interaction. The KYAPay drafts start from a related premise: a "chain of responsibility -- from agent action back to an accountable human" (Section 3.1 of [KYAPAY-USING]).¶
Audit and dispute resolution still need the identity, after the fact and for some requests only. Section 11 of [AIMS] requires audit events to record the "delegated subject (user or system), when present".¶
Today these two needs are met either by carrying a stable identifier everywhere, which is simple to audit and correlatable across services, or by pseudonymous identifiers whose resolution rests on the cooperation of the party that assigned them, under conditions that the specification does not state (Section 3.4.1 of [KYAPAY-TOKEN]; Section 5 of [KYAPAY-TOKEN]).¶
No specification states, as a set of checkable properties, what a principal reference reveals and to whom, who can undo the pseudonymity and under what condition, and how fast a withdrawal takes effect. This document addresses that gap. In one phrase, the aim is anonymity without loss of audit: a relying party can verify that an established principal stands behind the agent, proven and not named, and the principal is identified only under stated conditions.¶
It states six properties, each with a test that a reviewer can apply to a specification or a deployment (Section 3).¶
It maps each property to existing IETF text that provides part of it (Section 4).¶
It states an evaluation method (Section 5) and applies it to five drafts (Section 6).¶
It lists the security and privacy considerations that the properties raise (Section 8 and Section 9).¶
See Section 7. In short: no format, no protocol, no identity-proofing method, no trust framework, no ranking and no recommendation of an implementation.¶
The natural person or organization that answers for a request that an agent makes. Section 10.3 of [AIMS] calls this the "User or System"; Section 3.4.1 of [KYAPAY-TOKEN] calls it the "human principal".¶
What a request carries that refers to the principal: an identifier, a proof or a commitment. A principal reference is distinct from the agent's identifier.¶
The party that established the principal, by identity proofing or an equivalent process, and issued the principal reference. It is the party that can map a reference to the principal.¶
A party that receives a request with a principal reference and decides on it. This document says "relying party" because "verifier" means the identity-verification provider in Section 3.4.1 of [KYAPAY-TOKEN] and the recipient of a presentation in [RFC9901].¶
As used in OAuth identity chaining (Section 2.5 of [ID-CHAINING]): a set of parties under one authorization server's policy.¶
The sequence of requests from the first agent to the last relying party, possibly across trust domains.¶
Mapping a principal reference to the principal's identity.¶
The reference issuer's published statement of who may request resolution, on what grounds, what is returned and to whom, and what record is kept.¶
Whether a principal reference, or the principal's standing behind it, is still accepted. Principal standing and token status are different checks: "a status list revokes a token, not a principal" (Section 3.6 of [WIMSE-EVAL]).¶
The event from which a withdrawal bound is measured, for example the issuer's decision, the publication of the status list or commitment that carries it, or the principal's own request.¶
This document is Informational and places no requirement on any implementation. It uses no BCP 14 requirement keywords of its own.¶
Each property is stated and then followed by a test. A specification provides a property when every conforming implementation of it has the property; it allows the property when a deployment can choose to have it.¶
A relying party can verify, from what the request carries or references:¶
that a principal reference is bound to the request and to the acting agent;¶
which reference issuer issued it; and¶
the assurance at which the issuer states it established the principal.¶
The principal reference is distinct from the agent's identifier.¶
Remove the reference, or replace it with one bound to another agent or request: the relying party's check fails. Present a reference from an issuer outside the relying party's accepted set: the relying party refuses. The agent identifier and the principal reference are different values.¶
Every party in a delegation chain that receives a principal reference can establish that it refers to the same principal as the reference received by the preceding party. Within a trust domain this can be equality of the value. Across trust domains the value may change, and the party that changed it then records the correspondence (value received, value issued, its own identity) in a record that it signs.¶
Build a chain in which one hop carries a reference to a different principal. The receiving party detects it, or the signed records of the chain show the break.¶
No party other than the reference issuer can, from the principal references it receives and without the issuer's participation:¶
obtain an attribute that identifies the principal, beyond those the relying party requested and the principal authorized; or¶
link two references to the same principal presented in different trust domains.¶
Two references to the same principal, presented in two trust domains, carry no common value other than the issuer's identity and public parameters: no common identifier, signature, salt, key or status-list index. The specification lists any common value as a residual.¶
Item (b) is the cross-domain case of the distinct "sub" values that "cannot be correlated in the event of resource server collusion" (Section 6 of [RFC9068]) and of Verifier/Verifier unlinkability (Section 10.1 of [RFC9901]). P3 does not cover behavioral correlation (timing, IP addresses, the content of requests); see Section 9.¶
Only the reference issuer, or a party that the issuer names in its resolution process, can resolve a principal reference. The issuer publishes its resolution process, resolves only on a request that the process admits, and keeps a record of each resolution (reference, requester, ground, time, outcome) for the period that the process states.¶
The relying party can retrieve the issuer's resolution process before relying on a reference. An auditor can sample resolution records against requests. The record-keeping half can be checked by audit only; a specification can say no more than that the record exists and what it contains.¶
P4 is the property for which the least existing text was found (Section 4).¶
A relying party can establish the current status of a principal reference, and of the principal's standing where that is a separate check, without sending to the issuer or to a status provider any value that identifies which reference it is checking.¶
The status request carries no per-reference value: it fetches a list or a commitment shared by many references (herd privacy, Section 12.1 of [STATUS-LIST]), or the status reaches the relying party without a request.¶
P5 concerns status checks. A reference minted for one named audience tells the issuer the audience at issuance time. That is a separate observation; the evaluation records it where it applies (Section 6) and does not grade it as P5.¶
For each way in which a principal reference or a principal's standing can be withdrawn, the specification or the deployment states:¶
the withdrawal event from which the bound is measured; and¶
the maximum time after that event until a conforming relying party refuses requests that rely on it.¶
"Immediately" and "without undue delay" are not stated bounds.¶
Withdraw at time t0, measured from the stated event. A conforming relying party refuses at or before t0 plus the bound. Bounds measured from the receipt of a notification do not count, since the time of receipt is not observable by the party that withdrew.¶
For each property, this section names the existing text that provides part of it and what that text leaves open.¶
Existing text: "act" identifies the acting party separately from the subject (Section 4.1 of [RFC8693]); Section 10.3 of [AIMS] puts the agent in "client_id" and the principal in "sub". Left open: the assurance of the principal, and the issuer of the principal reference as distinct from the token issuer.¶
Existing text: "Transcribing the subject identifier", where the mapping can happen at the token exchange step or at the assertion step (Section 2.5 of [ID-CHAINING]). Left open: where the mapping is recorded, and that it stays out of the receiving domain.¶
Existing text: distinct "sub" values per resource server against collusion (Section 6 of [RFC9068]); "Data Minimization" in claims transcription (Section 2.5 of [ID-CHAINING]); the four kinds of unlinkability, batch issuance for Verifier/Verifier unlinkability, and the limit that salted-hash formats cannot give Issuer/Verifier unlinkability against a colluding verifier (Section 10.1 of [RFC9901]). Left open: non-linkage across trust domains for references passed along a chain.¶
Existing text: none that conditions resolution. The nearest phrases are audit records "retained according to the security policy of the deployment" (Section 11 of [AIMS]); issuer records that let an issuer "resolve a token, given its iss and jti, to the verifications and authorizations behind it", with their disclosure left to applicable law (Section 5 of [KYAPAY-TOKEN]); and the requirement that a WHO profile define "a hiding commitment construction, its scope, its binding to the verified WHO record, and how an authorized verifier obtains any required disclosure" (Section 6.2 of [ACCOUNTABILITY-COMPOSITION]). Left open: the condition on resolution and the record of it.¶
Existing text: herd privacy, under which the issuer "does not learn for which Referenced Token the Relying Party is requesting the Status List" (Section 12.1 of [STATUS-LIST]); issuer tracking through one list per token (Section 12.2 of [STATUS-LIST]); the list index as traceable data (Section 12.5 of [STATUS-LIST]); a revocation callback "could potentially disclose information about the credential's usage to the Issuer" (Section 10.1 of [RFC9901]). Left open: principal standing as distinct from token status.¶
Existing text: "exp" and "ttl" of a status list (Section 13.7 of [STATUS-LIST]; Section 11.5 of [STATUS-LIST]); principal standing established from a source the relying party pins, within a maximum age it pins, failing closed (Section 3.6 of [WIMSE-EVAL]). Left open: the start event of the bound, per withdrawal path.¶
Related, not mapped: Section 9 of [AUDIT-ARCH] observes that, at an organizational boundary, the Tracing Delegation Chain "could be replaced ... by an opaque identifier whose mapping is retained only at the originating side's Audit Store". That is P2 and P3 at the audit layer, with P4 left open.¶
The current revision of each draft as published on the Datatracker, read in full. Deployments, products, editors' copies and other revisions are not evaluated.¶
One of the following.¶
Provides: the text requires behavior that meets the property, as an absolute requirement or a recommendation in the terms of BCP 14.¶
Allows: the text permits it (an optional feature, an optional claim, or a choice left to deployment).¶
Partial: the text meets part of the property; the entry names the part.¶
Not addressed: the text says nothing that meets or prevents it.¶
Prevents: the text requires behavior that rules the property out.¶
Out of stated scope: the draft itself declares the topic out of its scope.¶
Every entry cites a section. Residuals that a draft states about itself are credited to the draft. A property that a draft does not aim at is still graded, and the notes for that draft say what it does aim at.¶
The five drafts solve different problems, and a draft that grades low here may be the right tool for its own purpose. This document does not rank them.¶
The authors of each evaluated draft are asked to correct the entries for their draft. Corrections are taken, or recorded with the reason they were not taken, in the next revision.¶
This first pass covers the following revisions, each confirmed as the current one on the Datatracker on 18 November 2026:¶
the KYAPay family: [KYAPAY-TOKEN], [KYAPAY-USING] and [KYAPAY-EXCHANGE] (16 November 2026), cited below as "token", "using" and "exchange";¶
[AGENT-PASSPORT], draft-le-agent-passport-00 (30 October 2026);¶
[DELEGATE-SD-JWT], draft-ietf-oauth-delegate-sd-jwt-00 (dated 26 October 2026, posted 30 October 2026);¶
[DELEGATOR-BOUND], draft-nandakumar-wimse-delegator-bound-chains-00 (16 November 2026).¶
The summary table gives the grades. The subsections that follow give, for each draft, the reasons and the sections they rest on. The table is a summary of those entries, not a score.¶
| Property | AIMS -01 | KYAPay family | Agent Passport -00 | Delegate SD-JWT -00 | Delegator-Bound -00 |
|---|---|---|---|---|---|
| P1 | Partial | Partial | Partial | Allows | Partial |
| P2 | Allows | Not addressed | Provides | Provides | Provides within a chain |
| P3 | Partial | Provides, with stated residuals | Prevents within a chain | Partial | Allows, with stated residuals |
| P4 | Not addressed | Partial | Not addressed | Not addressed | Not addressed |
| P5 | Not addressed | Not addressed | Partial | Not addressed | Partial |
| P6 | Partial | Partial | Partial | Partial | Partial |
The agent is conveyed in "client_id" and the principal in "sub" (Section 10.3 of [AIMS]). The text covers the authentication of a user who delegates (Section 10.5.1 of [AIMS]) and says nothing on how the principal was established or at what assurance.¶
Cross-domain access should use identity chaining (Section 10.7 of [AIMS]), which permits transcription of the subject identifier; where a transcribed value and its correspondence are recorded is not stated. Section 11 of [AIMS] states that end-to-end audit is enabled by stable, verifiable identifiers.¶
Minimization is required: user identity information delegated to agents is "not exposed to unrelated services", and cross-domain exchanges disclose only what the target authorization decision needs (Section 10.10 of [AIMS]). Nothing addresses non-linkage across trust domains, and the audit model of Section 11 of [AIMS] relies on stable identifiers.¶
Audit records are "retained according to the security policy of the deployment" (Section 11 of [AIMS]); nothing conditions the mapping of "sub" back to the user.¶
The status paths named are Shared Signals events that a participant may subscribe to (Section 11 of [AIMS]) and token introspection, a call to the authorization server per token (Section 10.3 of [AIMS]; Section 10.10 of [AIMS] prefers opaque tokens with introspection for claim minimization). The text does not consider what either path shows the party that answers.¶
Revoked authorization is enforced "without undue delay", and cached decisions and tokens are not used "after a revocation or risk notification is received" (Section 11 of [AIMS]): no number, and the bound is measured from receipt.¶
AIMS composes existing specifications (Section 14 of [AIMS]; Section 15 of [AIMS]) and leaves compliance criteria and the policy format to deployments (Section 12 of [AIMS]; Section 13 of [AIMS]). It is a natural place for a pointer to these properties. An open pull request on its Section 10.10 ([AIMS-PR176], not merged on 18 November 2026) would add subject identifiers specific to a party or trust domain, claims transcription with a retained mapping, and resolution "according to the security policy of the deployment". If merged as proposed, P3 would move to Allows and P4 to Partial. This evaluation grades -01 only.¶
Items (b) and (c) are provided. The human principal claim "hid", the agent platform claim "apd" and the agent claim "aid" are each required and distinct (token, Section 3.4), and "aid" "identifies the agent, not the human principal on whose behalf the agent acts" (token, Section 3.4.3); the issuer is "iss"; the methods by which the principal was verified and authenticated can be conveyed in "ivm", "amr" and "aml", which an issuer should include when it knows them, and the time of authentication in "auth_time" (token, Section 3.4.1); and a recipient derives assurance per layer (using, Section 6.2). Item (a) is a deployment choice: the token is bound to one Target (token, Section 3.1.1), but it is a bearer token unless it carries a confirmation key (token, Section 6.5; using, Section 4.1).¶
A token is minted for one Target ("aud", "tsi"; token, Sections 3.1.1 and 3.4) and is not forwarded beyond the parties needed for the interaction (token, Section 6.2). The exchange for an access token carries no actor token (exchange, Section 4). The "chain of responsibility" is the three identity layers within one token (token, Section 2.1.1). No text lets a Target pass the principal reference on to a further party.¶
"hid.id", "sub" and "aid.id" differ per Target, and "hid.id" also per agent platform; values for different platforms or Targets are not "linkable to one another without the issuer's cooperation" (token, Sections 3.4.1 and 3.4.3), and two Targets "cannot link the principal from these values" (token, Section 7). Directly identifying claims are included only when a Target requires them for the action at hand and the principal has authorized their disclosure (token, Sections 3.4.1 and 7). Residuals stated by the draft: "apd" and "aid.name" are the same at every Target and can link a principal's activity (token, Section 7); "verification_id" is the same at every Target and is disclosed under the same condition as the identifying claims (token, Section 3.4.1); and "whether the services of one Target share a value is left to the issuer" (token, Section 7).¶
The part provided is the record that makes resolution possible. An issuer "MUST keep the following records": verification records, authorization records for each disclosure that needs the principal's authorization, and issuance records that tie each token to the verifications and authorizations it relied on; together they let an issuer "resolve a token, given its iss and jti, to the verifications and authorizations behind it", and the issuer must protect them against unauthorized access (token, Section 5). The condition is not stated: the specification "does not set how long an issuer keeps these records, or when it discloses them to third parties such as courts or regulators", which it leaves to applicable law, and it "does not define an interface for querying them" (token, Section 5). The text requires neither a published resolution process nor a record of each resolution. The feedback channel from a recipient to the issuer is left open (using, Section 6.10).¶
No revocation mechanism is defined; recipients should consult "an issuer's live-status or revocation endpoint, where one is offered" (using, Section 8.8), which is not specified. Separate observation: a token minted per Target tells the issuer the Target at issuance, and the issuance records keep "aud" and "tsi" for each token (token, Sections 3.1.1, 3.4 and 5); the privacy considerations cover Targets and third-party observers (token, Section 7).¶
There is no revocation (using, Section 8.8). The only bound is the token lifetime: issuers set "exp" "to the shortest value compatible with the intended interaction" (token, Section 6.4), and recipients reject tokens whose lifetime exceeds their local policy maximum (using, Section 8.2). Both are measured from issuance. A token attests authorization "at the time of issuance" (using, Section 8.9).¶
The KYAPay family starts from the premise of this document (Section 1.1), and it is the only one of the five that provides P3, through pairwise identifiers, and the only one that requires the issuer records on which resolution rests. Its documents state their own boundaries: no revocation (using, Section 8.8), issuer trust "an open problem analogous to the Certificate Authority model" (using, Section 8.7), bearer use by default with an optional confirmation key (token, Section 6.5; using, Section 4), and retention and disclosure of issuer records left to the law that applies to the issuer (token, Section 5). Two further readings bear on P4. "Where a separate verifier performed the verification, the issuer keeps the verifier's identity, the verification_id and the result, and the verifier keeps the rest of the record" (token, Section 5): P4's "issuer, or a party the issuer names" then covers a pair of parties, and neither is given a stated process. Keyed derivation of "hid.id" (token, Section 3.4.1) changes how the issuer computes a mapping, not what a resolution record under P4 contains. The companion drafts on verification methods, authentication method values and AML/CFT methods define claim values and bear on P1, item (c), only. The issuer records come from a pull request merged on 15 November 2026 ([KYAPAY-PR72]). An open issue plans exchange for a sub-agent's token at the issuer, with the issuer keeping "a record linking the issued token to the subject token" ([KYAPAY-EXCHANGE-ISSUE10]); it bears on P2. This evaluation grades the published revisions only.¶
The accountable party is the controller's origin (Section 1.2 of [AGENT-PASSPORT]; "agent_controller", Section 4.2 of [AGENT-PASSPORT]; "contact", Section 4.5 of [AGENT-PASSPORT]). The principal is the root receipt's "sub", "as identified by the root receipt's authorisation server" (Section 5.2 of [AGENT-PASSPORT]), with no stated assurance. Binding is provided: the leaf receipt is bound to the key that signs the request (Section 5.5 of [AGENT-PASSPORT], step 4).¶
"Every receipt in a chain" carries "the same sub as the root" (Section 5.2 of [AGENT-PASSPORT]), checked at each step (Section 5.5 of [AGENT-PASSPORT]).¶
"Receipts identify the principal (sub) to every RP they are shown to" (Section 10 of [AGENT-PASSPORT]), and the same-"sub" rule (Section 5.2 of [AGENT-PASSPORT]) excludes values specific to a trust domain. Linkage of the agent across relying parties can be avoided with distinct holder keys and passports per relying party (Section 10 of [AGENT-PASSPORT]).¶
The receipt carries the principal's identifier as the root authorization server assigns it (Section 5.2 of [AGENT-PASSPORT]); nothing addresses mapping it back or recording that mapping (Section 10 of [AGENT-PASSPORT]).¶
Passports and receipts should carry a Token Status List reference (Section 4.2 of [AGENT-PASSPORT]; Section 5.2 of [AGENT-PASSPORT]; Section 7 of [AGENT-PASSPORT]), which gives herd privacy where lists are shared (Section 12.1 of [STATUS-LIST]; Section 12.2 of [STATUS-LIST]). There is no separate check of the principal's standing. Separate observation: each verification fetches the controller's "/.well-known/agent-controller" document, cacheable for up to 24 hours (Section 4.5 of [AGENT-PASSPORT]).¶
Stated: a controller's withdrawal of authorization takes effect within the 24-hour cache of its document (Section 4.5 of [AGENT-PASSPORT]); a passport lives at most 86400 seconds (Section 4.2 of [AGENT-PASSPORT]); status lists are cached for no longer than their time-to-live (Section 7 of [AGENT-PASSPORT]). No bound is stated for the principal's own standing.¶
Accountable organizations and verifiable builds, with attenuation guaranteed by construction (Section 1.2 of [AGENT-PASSPORT]). The low P3 and P4 grades follow from that aim: the controller is meant to be known. Its open issues include a root receipt signed directly by the principal's authenticator (Appendix B of [AGENT-PASSPORT]).¶
The properties of the principal depend on the credential delegated; the verifier can "link the presentations back to the initial Issuer" (Section 4 of [DELEGATE-SD-JWT]). Nothing addresses the principal's assurance.¶
Each Key Binding SD-JWT in the chain is bound to the preceding one (Section 4 of [DELEGATE-SD-JWT]; Section 6 of [DELEGATE-SD-JWT]; Section 8.1 of [DELEGATE-SD-JWT]).¶
Selective disclosure applies to both the original credential and the delegate payload (Section 4.1 of [DELEGATE-SD-JWT]), and the privacy considerations of Section 10 of [RFC9901] apply (Section 9 of [DELEGATE-SD-JWT]). By Section 10.1 of [RFC9901], presentations of one credential are linkable across verifiers unless batch-issued, and the verifier sees the full chain (Section 4 of [DELEGATE-SD-JWT]).¶
Beyond the privacy considerations inherited from [RFC9901] (Section 9 of [DELEGATE-SD-JWT]), nothing addresses resolution.¶
Not addressed for delegations: a holder "can not easily distribute revocation information to Verifiers" (Section 8.3 of [DELEGATE-SD-JWT]). The initial SD-JWT inherits Section 10.1 of [RFC9901].¶
A short "exp" and claims constraining the use of the delegated SD-JWT "limit this problem" (Section 8.3 of [DELEGATE-SD-JWT]); no bound is stated.¶
Onward presentation of an SD-JWT to a delegate with minimal disclosure (Section 1 of [DELEGATE-SD-JWT]). Principal properties come from the credential delegated. Its note on revocation (Section 8.3 of [DELEGATE-SD-JWT]) is the P6 gap as stated by its authors.¶
The draft extends Delegation Tokens ([DELEGATION-CHAIN]), which this revision does not evaluate (Appendix A). The entries grade the extension's own text, and the base only where the extension states what it takes from it.¶
Items (a) and (b) are provided in a delegator-bound chain. The principal is carried in "sub", distinct from the acting agent in "client_id" (Section 2 of [DELEGATOR-BOUND]). Each child is signed with the key that its parent confirms, so a verifier knows that "each child was issued by the holder of the key that its parent confirmed, and therefore by the Acting Agent that the parent named" (Section 3.4 of [DELEGATOR-BOUND]); the same key is the one bound for proof of possession (Section 3.4 of [DELEGATOR-BOUND]), and the last agent proves possession of the key that the last token confirms (Section 2 of [DELEGATOR-BOUND]). An optional "ctx" binds a chain to the dialog or transaction it was issued for (Section 7 of [DELEGATOR-BOUND]). The root token is verified under a trusted root key (Section 3.3 of [DELEGATOR-BOUND]), and how the verifier trusts that key "is outside this document" (Section 10 of [DELEGATOR-BOUND]). Nothing addresses how the principal was established or at what assurance.¶
A child follows "the principal invariant" of the base (Section 3.2 of [DELEGATOR-BOUND]); the verifier checks at each hop "that sub, sub_kind, and del_bind are unchanged" (Section 2 of [DELEGATOR-BOUND]) and rejects a chain in which a child omits or changes "sub_kind" (Section 4 of [DELEGATOR-BOUND]). Continuity is by equality, also between organizations whose agents delegate directly (Section 9 of [DELEGATOR-BOUND]). A chain that crosses a trust domain through an authorization server is re-issued there (Section 6 of [DELEGATOR-BOUND]), or the delegator at the boundary obtains a new root, "ending the chain" (Section 11 of [DELEGATOR-BOUND]); whether the new chain refers to the same principal, and where that correspondence is recorded, is not stated.¶
When the issuer of the root token marks "sub" as a pseudonym, "sub" is a keyed hash of the issuer's identifier for the principal and the root token's "jti", and "The pseudonym differs for every chain, so chains for the same Principal cannot be linked by sub" (Section 4 of [DELEGATOR-BOUND]; Section 11 of [DELEGATOR-BOUND]). Within one chain every agent and enforcement point sees the same value, which is what P2 needs. An issuer must not place in "sub" an identifier that another issuer directed to a specific audience (Section 4 of [DELEGATOR-BOUND]). The pseudonym is optional: "The value identifier, or absence, means that sub identifies the Principal directly" (Section 4 of [DELEGATOR-BOUND]). Residuals stated by the draft: "A delegation chain discloses the client_id of every hop to each later hop and to the enforcement point", and per-chain "client_id" values reduce that linkability at the cost of cycle detection (Section 11 of [DELEGATOR-BOUND]; Section 5 of [DELEGATOR-BOUND]); "Anyone who holds K can link pseudonyms to Principals" (Section 10 of [DELEGATOR-BOUND]).¶
"The issuer of DT_0 can resolve the pseudonym, for example for audit or revocation" (Section 4 of [DELEGATOR-BOUND]). Nothing states when it resolves, to whom it returns the result, or what record it keeps. The pseudonym is a keyed hash, so the issuer resolves it from a stored record of the principal for which each root token was issued, or by computing the value for each of its principals, and that record is sensitive data about principals. The issuer must protect the key and should rotate it (Section 10 of [DELEGATOR-BOUND]), which protects the ability to resolve without conditioning it.¶
"Short lifetimes and status lists" limit the exposure of a compromised key, and a verifier should deny a chain when it cannot obtain current status for a token that carries a status claim (Section 10 of [DELEGATOR-BOUND]). Where that status is a Token Status List shared by many tokens, the check has herd privacy (Section 12.1 of [STATUS-LIST]). Status is per token; there is no separate check of the principal's standing.¶
"Revoking DT_i invalidates every chain through it", and a verifier should deny a chain whose status it cannot obtain "within its configured staleness bound" (Section 10 of [DELEGATOR-BOUND]). The bound is each verifier's configuration, not a time measured from the issuer's decision, and no value is stated. Withdrawing the principal, rather than one chain, means revoking every root token issued for it, since relying parties cannot relate the pseudonyms; that path and its bound are not stated.¶
Evidence at every hop that the agent before it granted the hop, with no authorization server after the root (Section 3.4 of [DELEGATOR-BOUND]): "This extension is for agents that delegate directly to each other" (Section 9 of [DELEGATOR-BOUND]). It is the only one of the five that keeps the principal reference equal along a chain and different between chains, and it maps the same rules onto Delegate SD-JWT (Section 8 of [DELEGATOR-BOUND]). Its resolution sentence (Section 4 of [DELEGATOR-BOUND]) is where a stated resolution process would attach.¶
No draft provides P1 in full. The KYAPay family states the issuer and the assurance and leaves the binding to the agent optional; Agent Passport and the delegator-bound chains bind the reference to the agent's key and state no assurance for the principal.¶
P2 is provided by Agent Passport and Delegate SD-JWT, and by the delegator-bound chains within a chain.¶
P3 is met by the pairwise identifiers of the KYAPay family, within its single-hop scope, and allowed by the per-chain pseudonym of the delegator-bound chains, which keeps one value along a chain and a different one for each chain. The other drafts that keep the principal constant along a chain do it by identifier equality or by showing the full chain, which works against P3 across trust domains.¶
P4 is met in part by the KYAPay family only: its issuer records make resolution possible, and the condition on disclosure is left to applicable law. No draft requires a published resolution process or a record of each resolution. The delegator-bound chains state that the issuer of the root token can resolve the pseudonym, and leave the condition open.¶
P5 and P6 are met in part, and P5 only where a draft adopts Token Status List. None of the five bounds the principal's own standing from a stated event.¶
Token formats, claim names, media types, protocols and wire encodings.¶
Identity-proofing methods and assurance scales.¶
How a relying party decides which reference issuers to accept (trust frameworks).¶
The legal basis for resolution, and who may compel it.¶
Cryptographic constructions. Pairwise identifiers, claims transcription, commitments and anonymous credentials can each meet some of the properties; none is recommended here.¶
Rankings, conformance claims and recommendations of implementations.¶
P3 and P4 together put the ability to identify the principal at one party. Compromise or coercion of that party exposes every principal it issued for. [AIMS-PR176] says the same of transcription: it "concentrates the ability to identify the User at that server". A deployment can reduce this by splitting the ability to resolve between parties that must act together; the properties neither require nor exclude it.¶
Under P2 across trust domains, a relying party relies on the transcribing party's statement that two values refer to one principal. A false statement is not detectable from the reference alone; the signed record makes it attributable.¶
A principal reference that is not bound to the request and to the agent (P1, item (a)) can be replayed by another agent. The tokens of [KYAPAY-TOKEN] are bearer tokens unless they carry a confirmation key (Section 6.5 of [KYAPAY-TOKEN]).¶
Where status or standing cannot be established within the stated bound, the properties assume refusal, as in Section 3.6 of [WIMSE-EVAL]. A specification that fails open in that case does not meet P6.¶
The resolution process is itself an attack surface: false requests, overbroad grounds, social engineering of the issuer. The record that P4 requires is what makes such abuse auditable.¶
P5 depends on the number of references per list and on the issuer not issuing one list per token (Section 12.1 of [STATUS-LIST]; Section 12.2 of [STATUS-LIST]).¶
A reference minted for a named audience tells the issuer the audience. This is outside P5 as stated and is to be weighed separately.¶
IP addresses (Section 12.1 of [STATUS-LIST]), request timing (Section 10.1 of [RFC9901]), stable agent names (Section 7 of [KYAPAY-TOKEN]) and the content of requests can link a principal across trust domains whatever the reference does.¶
The record that P4 requires is sensitive data about principals. Its retention period and who may access it belong in the stated resolution process.¶
P4 limits resolution to stated grounds. Whether those grounds are acceptable is a policy question outside this document.¶
This document has no IANA actions.¶
The following drafts address agent delegation or agent identity and are candidates for a later revision. They are listed here so that their absence from Section 6 is not read as a judgment.¶
[AUTHZ-INFO-SHARING], draft-diaconu-agents-authz-info-sharing-01;¶
[DELEGATION-CHAIN], draft-asor-wimse-agent-delegation-chain-03;¶
[CREDENTIAL-DELEGATION], draft-sweeney-wimse-credential-delegation-02;¶
[AGENT-MANDATE], draft-hulea-wimse-agent-mandate-01;¶
[ATTENUATING-TOKENS], draft-niyikiza-oauth-attenuating-agent-tokens-02;¶
[AGENT-SD-CARD], draft-nandakumar-agent-sd-card-00.¶