Skip to content

Spec requirements checklist (as of 2026-11-18)

Target revisions, checked on datatracker: txn-tokens -11 (2026-07-30, awaiting WG write-up), AIMS -01 (2026-11-02, Informational, WG doc), ID-JAG -06 (2026-10-31), identity-chaining -17 (2026-07-19, RFC Ed Queue), AuthZEN Authorization API 1.0 Final (OIDF approval announcement dated 2026-01-12).

For txn-tokens, ID-JAG and chaining, I read the published draft text (-11, -06, -17) and confirmed it matches the latest revision on datatracker.

1. Transaction Tokens (draft-ietf-oauth-transaction-tokens-11)

Section titled “1. Transaction Tokens (draft-ietf-oauth-transaction-tokens-11)”
  • Header typ MUST be txntoken+jwt (§9.1). kid SHOULD be present
  • REQUIRED claims: iat, aud, exp, txn (RFC 8417 §2.2), sub, scope, req_wl (§9.2)
    • aud MUST identify the Trust Domain. Other domains MUST NOT accept the token
    • sub MUST be unique within the aud domain. It is not tied to iss
  • RECOMMENDED claims: tctx (immutable authorization details) and rctx (environment context, e.g. req_ip, authn, transport)
    • In -11 the claim is rctx. There is no req_ctx claim. The only remaining req_ names are req_wl and the request parameters request_context / request_details
  • iss is OPTIONAL. It identifies the issuer when the signing key is not known in advance (§13.16)
  • Recipients MUST ignore claims they do not understand (§9.2)
  • Requests to a remote TTS over HTTP MUST use RFC 8693 (§5.1, §11)
  • Required parameters:
    • grant_type=urn:ietf:params:oauth:grant-type:token-exchange
    • audience: the Trust Domain name
    • scope: the purpose
    • requested_token_type=urn:ietf:params:oauth:token-type:txn_token
    • subject_token and subject_token_type
  • RECOMMENDED parameters: request_context (JSON → rctx) and request_details (JSON → tctx), sent form-urlencoded
  • subject_token_type (§11.2):
    • Any type from RFC 8693 §3 is allowed except refresh_token
    • urn:ietf:params:oauth:token-type:self_signed (§11.2.1): a JWT that MUST contain iss = requesting workload, sub, aud = TTS ID (the TTS SHALL check it), iat, exp
    • urn:ietf:params:oauth:token-type:unsigned_json (§11.2.2): sub MUST be present
    • Custom URNs are allowed
  • Processing (§11.3), all MUST:
    • Client authn and authorization
    • Validation of the subject_token, including the signature if signed
    • Uniqueness of sub
    • Setting iat, aud, exp, txn
    • Evaluation of scope
  • MUST NOT issue from an expired subject_token (§6)
  • Response (§11.4): token_type=N_A, access_token = the Txn-Token, issued_token_type=...:txn_token. refresh_token MUST NOT be returned. Errors use the RFC 6749 §5.2 format
  • Mutual authentication MUST be used (§11.5). Asymmetric methods (RFC 7523, RFC 8705, WIMSE mTLS, HTTP-sig, WPT) SHOULD be used
  • Exactly one logical TTS per Trust Domain (MUST, §10)
  • Header Txn-Token: (MUST), carrying exactly one token. The Authorization header MUST NOT be used
  • The recipient MUST validate three things: the signature, that aud is its own domain, and that the token has not expired (§12.2)
  • Outbound calls MUST carry the received token unmodified
  • The lifetime MUST be short (§13.1). Replay is detected with txn. A single-use check is optional (§13.2)
  • A Txn-Token MUST NOT embed an access token (§13.4)
  • scope MUST NOT exceed the subject_token’s scope (§13.6). If the subject_token’s scope is unknown, the request MUST be rejected (§13.14)
  • TTS configuration MUST come from a trusted source (§13.7). The TTS MUST be authenticated (§13.9)
  • A Txn-Token MUST NOT be used as an authn credential (§13.12) or as an OAuth AT (§13.13)
  • Replacement (§13.15), when a Txn-Token is passed as the subject_token:
    • Narrowing and adding values MAY happen. Broadening MUST NOT
    • txn, sub, aud MUST NOT change
    • An expired token MUST NOT be replaced
    • The lifetime MAY be extended. If it is, the number of replacements SHOULD be limited
    • The call chain MUST be preserved, but the method is out of scope
  • A Txn-Token MUST NOT be logged in full (§15.2). Log a hash or the payload without the signature
  • Identity-chaining SHOULD be used across domains (§14)
  • -11: history typo fix only
  • -10: editorial changes to claim descriptions
  • -09 (changes from -08):
    • Added a TTS issuance policy section
    • Added cross-domain guidance
    • Added token revocation (introspection) to the considerations
    • Removed the rctx-specific claim definitions and renamed the section to Request Context
    • “Clarified” how req_wl represents multiple workloads
    • Changed request parameters and body claims from OPTIONAL to RECOMMENDED
  • -08: made iss always OPTIONAL
  • Earlier:
    • Renamed purpose to scope
    • Renamed azd to tctx
    • Removed the replacement text from the body and consolidated it into §13.15
  • In the §9.2 claim list, tctx points to §9.2.2 and rctx to §9.2.3, but the sections are the other way round
  • req_wl:
    • The history says the multiple-workload representation was clarified, but the body only says “A string value”
    • The IANA registration (§16.2) lists only tctx and rctx, not req_wl
  • The method for preserving the call chain on replacement is undefined. visado has to choose between concatenating req_wl and its own claim
  • The audience parameter (Trust Domain) and the self-signed JWT aud (TTS ID) are different values
  • How to compare scopes for “scope ≤ subject_token’s scope” is undefined. §9.2.1 allows internal scopes to mean something different from external ones, which sits awkwardly with that rule
  • No type URN exists for a WIT or WPT as the subject_token, so it would be a custom URN

2. AIMS (draft-ietf-wimse-aims-01, Informational)

Section titled “2. AIMS (draft-ietf-wimse-aims-01, Informational)”
  • 8 layers: Identifier → Credentials → Provisioning → Authentication → Authorization → Observability/Remediation → Policy → Compliance
  • A conceptual model, not a protocol
  • Roles:
    • Agent (= workload, §4)
    • LLM
    • Tools/Services/Resources
    • User/System (+ other Agents)
    • IdP, AS/RS, TTS, external PDP (§10.4)
    • posture verifier (§8)
    • SSF receivers and monitoring systems (§11)
AIMS § flow specs combined
§6 identifier WIMSE ID. Exactly one per agent (MUST). May be a SPIFFE ID (MAY)
§7–8 credential WIT/WIC (WIMSE-CRED), X.509-SVID / JWT-SVID / WIT-SVID. Short-lived (SHOULD), exp required (MUST). LLM access to credentials MUST NOT be allowed
§9.1 / §9.2 authentication mTLS / WPT (wth) or WIMSE HTTP-sig (RFC 9421). Relay and replay considerations are MUST
§9.3 long-running session Re-validate on expiry. WPT per message
§10.2–10.5 OAuth authorization code and client credentials. Client authn via RFC 7523 / RFC 8705 / OAuth-SPIFFE. Users use a phishing-resistant method (SHOULD)
§10.4 policy AuthZEN. The token is one input to the decision. Sender-constrained (mTLS or DPoP)
§10.6 / §10.9 internal propagation Txn-Token and RFC 8693. Forwarding the AT is an anti-pattern
§10.7 cross-domain identity-chaining (SHOULD), ID-JAG, Txn-Token assertion grant
§10.8 HITL CIBA. A local UI confirmation alone MUST NOT count as authorization. Bind to the AS grant (MUST)
§10.11 discovery RFC 8414, RFC 9728, CIMD, RFC 7591
§11 monitoring and revocation SSF / CAEP / RISC / WISE. Enforce on receiving a revocation (MUST). An invalidated cache MUST NOT be used
  • Uppercase MUST/SHOULD appear in §6, §7, §8, §9.2.3, §10.x, §11, §13
  • §1–5, §9.1, §9.3, §14–16 are descriptive only
  • The whole document is Informational, so it reads as a recommended architecture, not as conformance requirements
  • Audit log: durable and tamper-evident (MUST)
  • Minimum fields (MUST):
    • authenticated agent ID
    • delegating subject
    • resource/tool
    • action and decision
    • timestamp and correlation ID
    • posture/risk state
    • remediation/revocation and its cause
  • Reconstruction of the execution chain (SHOULD)
  • Log tokens by ID or hash, not in full (§10.10)

PR #176 (ietf-wg-wimse/draft-ietf-wimse-aims)

Section titled “PR #176 (ietf-wg-wimse/draft-ietf-wimse-aims)”
  • Adds 5 paragraphs on “correlation and user identifiers in delegation chains” to §10.10 Privacy
  • Status: open (created 2026-11-09, updated 11-17; reportedly to be settled at the 11-18 session)
  • Points:
    • Claims transcription (chaining §2.5) may emit a different subject ID outside the domain
    • The mapping is kept in the audit record of the AS that did the transcription, not in the token
    • The subject ID is the same across all hops within a domain and may change only when crossing domains
  • The PR says it adds “no new normative text”, but it contains a MUST sentence (the same as an existing §10.10 sentence)
  • Wording is not final
  • It defines neither the act claim nor agent→agent delegation chains
  • visado takes delegation chain validation from a separate draft (e.g. asor agent-delegation-chain)

What the reference implementation should demonstrate (inference)

Section titled “What the reference implementation should demonstrate (inference)”
  1. The agent ID is a single WIMSE/SPIFFE ID, verifiable via WIT+WPT, HTTP-sig or mTLS
  2. An external AT or an ID-JAG-derived AT is exchanged for a Txn-Token and propagated downstream in the Txn-Token header (the AT is not forwarded)
  3. scope only narrows, for both Txn-Token and chaining
  4. The AuthZEN PDP is queried with subject=agent (the delegating user goes in properties or context), and a deny stops the request
  5. Cross-domain uses ID-JAG / chaining
  6. A tamper-evident audit record (e.g. a hash chain) holds the 7 items from §11, with txn as the correlation ID
  7. (Optional) Caches are discarded on a CAEP revocation, and step-up uses CIBA

3. ID-JAG (draft-ietf-oauth-identity-assertion-authz-grant-06)

Section titled “3. ID-JAG (draft-ietf-oauth-identity-assertion-authz-grant-06)”
  • requested_token_type=urn:ietf:params:oauth:token-type:id-jag
  • audience: the Resource AS issuer, exactly one (multiple → invalid_request, unknown → invalid_target, §4.3.3)
  • resource: RECOMMENDED. If obtained from PRM, it MUST be sent unmodified
  • scope and authorization_details are optional. scope has RS-side meaning, not OIDC scope meaning
  • subject_token: the IdP MUST accept an ID Token (...:id_token) or SAML (...:saml2) and MAY accept an RT (...:refresh_token)
  • actor_token is accepted, but no processing rules are defined (§9.10)
  • IdP validation:
    • Client authn MUST succeed, otherwise invalid_client
    • The subject_token is actually of the stated type
    • The ID Token aud matches the client_id
    • Reject if typ is a different kind (e.g. at+jwt)
    • The granted resource / scope / authorization_details MUST go into claims
  • Response: issued_token_type=...:id-jag, token_type=N_A. If scope differs from the request, return scope
  • Header typ MUST be oauth-id-jag+jwt
  • REQUIRED: iss, sub, aud (the RS-AS issuer), client_id (the client ID at the RS-AS), jti, exp, iat
  • OPTIONAL: resource, scope, authorization_details, act, tenant, auth_time, acr, amr, aud_tenant, aud_sub, email, sub_id, cnf
  • A multi-tenant issuer MUST include tenant (§6.1)
  • sub MUST be the same for the same RS-AS regardless of client (§5.2)
  • grant_type=urn:ietf:params:oauth:grant-type:jwt-bearer, or ...:jwt-dpop when DPoP-bound
  • assertion = the ID-JAG
  • A request with neither client authn nor sender-constraint MUST be rejected with invalid_client
  • Validation (§4.4.1), in addition to RFC 7521 §5.2:
    • typ
    • trusted iss
    • aud is its own issuer (if an array, exactly one element)
    • client_id MUST match the authenticated client
    • If multi-tenant, check tenant
    • If cnf is present, check PoP: DPoP uses jkt, mTLS uses x5t#S256. Reject if unsupported
    • resource is under its own jurisdiction
  • Metadata:
    • RS-AS: list urn:ietf:params:oauth:grant-profile:id-jag in authorization_grant_profiles_supported, and then grant_types_supported MUST include jwt-bearer
    • IdP: list id-jag in identity_chaining_requested_token_types_supported. If id_jag_signing_alg_values_supported is published, it MUST include RS256
  • The same ID-JAG MUST NOT be reused at a different downstream RS-AS (§9.4)
  • Changes in -06: subject token validation errors changed from invalid_grant to invalid_request, and the draft explains why access_denied is used
  • Changes in -05: added jwt-dpop, mTLS binding, single audience, invalid_target, and more

Extension io.modelcontextprotocol/enterprise-managed-authorization, spec 2026-07-28.

  • It references ID-JAG -04
  • Additional requirements:
    • audience MUST be the MCP AS issuer
    • If resource is included, it MUST be the MCP Server’s RFC 9728 resource ID and MUST also go in the ID-JAG resource claim
    • The AT MUST be audience-restricted to the MCP Server
    • Discovery uses authorization_grant_profiles_supported
    • An unregistered client may use a CIMD URL as its client_id
    • The client declares the extension as a capability
    • The client must not redirect to the MCP AS authorization endpoint
  • No mention of DPoP
  • The -06 changes (error codes, jwt-dpop) are not reflected on the MCP side

4. Identity Chaining (draft-ietf-oauth-identity-chaining-17, RFC Ed Queue)

Section titled “4. Identity Chaining (draft-ietf-oauth-identity-chaining-17, RFC Ed Queue)”
  • Token exchange at the A-AS: either resource (the B-AS URI) or audience (a logical name) is REQUIRED. Unknown values, or values policy does not allow, MUST be rejected per RFC 8693 §2.2.2 (§2.3.1–2)
  • The aud of the returned grant MUST identify the B-AS. A single aud is RECOMMENDED (§2.3.3)
  • The B-AS MUST validate per RFC 7523 §3, §3.1, and MUST match aud against its token endpoint or issuer. If the subject cannot be identified, it MUST reject (§2.4.2)
  • The B-AS SHOULD NOT issue RTs (§5.4)
  • Replay mitigations to evaluate: short lifetimes, single use, client authn (§5.5)
  • Metadata: identity_chaining_requested_token_types_supported (§3)
  • Claims transcription (§2.5): either AS can add, change or remove claims. The representation format is undefined. No scope escalation
  • Evaluation POST /access/v1/evaluation (§6)
    • Request:
      • subject{type,id}: REQUIRED
      • resource{type,id}: REQUIRED
      • action{name}: REQUIRED
      • properties and context: optional
    • Response: {decision:bool, context?}
    • On false, the PEP MUST NOT proceed
    • A deny is HTTP 200, not an error
  • Evaluations /access/v1/evaluations (§7)
    • Top-level values are defaults that item values override
    • options.evaluations_semantic: execute_all (default), deny_on_first_deny, permit_on_first_permit
    • Responses keep the request order. Per-item errors come back as decision=false with context
  • Search /access/v1/search/{subject|resource|action} (§8)
    • Omit the id of the search target. If sent, it MUST be ignored
    • Pagination: page.token / page.limit
    • Requests for later pages MUST be the same as the previous one
    • The end is next_token=""
  • Metadata GET /.well-known/authzen-configuration (§9)
    • REQUIRED: policy_decision_point and access_evaluation_endpoint
    • If policy_decision_point does not match the URL used to fetch it, the metadata MUST NOT be used
    • The endpoint values from metadata MUST be used
    • signed_metadata is optional
  • Transport (§10, §11)
    • HTTPS POST, JSON object
    • X-Request-ID is RECOMMENDED, and the PDP MUST echo it
    • TLS MUST be used
    • PEP authentication SHOULD be used (mTLS / OAuth / API key)
    • Unknown fields MUST be ignored
    • Errors: 400 / 401 / 403 / 500
  • How to carry agent delegation (user→agent) is undefined. The subject could be the agent or the user, and the delegation could go in subject.properties or context
    • The WG draft MCP Tool Authorization profile and AARP (adopted 2026-06) are worth following
  • When decision=true comes with context the PEP does not understand, the PEP MAY deny. visado has to pick a behavior
  • No type URN exists for a WIT/WPT or an ID-JAG-derived AT as a Txn-Token subject_token
  • AIMS does not specify delegation chains (nested act)
  • ID-JAG’s act has no processing rules either
  • Where to keep the subject mapping for audit is still under discussion in chaining transcription and AIMS PR #176
  • MCP EMA (based on -04) and ID-JAG -06 diverge on error codes and DPoP handling
  • AIMS Appendix A has history for -02 / -03, but datatracker only has -00 and -01 (presumably carried over from the predecessor klrc’s history)