Review · November 16, 2026
Identity Assertion JWT Authorization Grant (ID-JAG)
The Problem
Two apps both use the same IdP for SSO. One wants to call the other's API on behalf of the user. Today that takes an interactive OAuth flow between the two apps, and the IdP never sees it.
This draft lets the IdP broker that access. The client trades its ID Token for a short-lived grant, the ID-JAG, and redeems it at the other app's authorization server. No redirect and no consent screen. The draft calls the pattern Cross-App Access (XAA).
How It Fits In
This is an adopted working group draft, not an individual submission. It is at -06, on the Standards Track, and it profiles Identity Chaining, another working group draft.
There are no new endpoints or grant types. The flow is Token Exchange at the IdP followed by the JWT bearer grant at the Resource Authorization Server. The new pieces are a token type, a JWT typ, two claims (resource, aud_tenant), two authorization server metadata parameters, one client metadata parameter, and a saml-nameid subject identifier format (Section 11).
It does add one relationship. The Resource Authorization Server already trusts the IdP for login. Now it also trusts the IdP's grant as "input to delegated authorization" and skips user consent (Section 4.1).
The dependency list is long. I count eleven OAuth RFCs in the normative references, plus the JWT RFCs, OpenID Connect, and SAML.
So what is left before publication?
- A "TBD" paragraph is still in Section 9.3 on step-up. It was there in -04 too.
- Three normative references are unfinished: Identity Chaining, the individual draft draft-parecki-oauth-jwt-dpop-grant, and "OpenID Connect Enterprise Extensions 1.0 - draft 01" for the
tenantandaud_subclaims. - RFC 10070 updated RFC 7523 this month. The draft does not cite it.
- The media type registration in Section 11.1 is one paragraph with no registration template.
Summary
There are four steps (Section 4.1).
- The user signs in to the client through the IdP. The client gets an ID Token or SAML assertion.
- The client sends that to the IdP's token endpoint as a Token Exchange
subject_token. It names the target authorization server inaudienceand the API inresource. The IdP checks admin policy and returns an ID-JAG. - The client presents the ID-JAG to the Resource Authorization Server and gets an access token.
- The client calls the API.
The ID-JAG is a JWT with typ of oauth-id-jag+jwt. It names the user (sub), the target server (aud), and the client as registered at that server (client_id). The Resource Authorization Server still applies its own policy and issues its own tokens.
The IdP MUST authenticate the client (Section 4.3.3). The Resource Authorization Server SHOULD, and MUST reject a request that has neither client authentication nor a key-bound ID-JAG (Section 4.4).
What changed since -04
The docs folder has -04 and -06. Almost all of the change landed in -05. The draft grew from 65 pages to 80.
- mTLS binding of the ID-JAG, next to DPoP (Section 9.11.2).
- A key-bound ID-JAG can now replace client authentication at the Resource Authorization Server. In -04 the text was "SHOULD only be supported for confidential clients."
- DPoP-bound ID-JAGs now use the
jwt-dpopgrant type. A server that doesn't support thecnfmethod MUST reject (Section 9.11). resourcehandling got real rules: string comparison,invalid_target, and the access token MUST be restricted to the granted resources.- New security text on metadata validation, with a concrete relay attack (Section 9.5), and on subject token storage (Section 9.2).
- Privacy Considerations and an
id_jag_signing_alg_values_supportedparameter. - -06 itself is small. A bad subject token now returns
invalid_requestinstead ofinvalid_grant, to match RFC 8693.
Interoperability
The two token requests are tightly specified. Two implementations built from this text will agree on the wire format. RS256 is mandatory on both sides (Section 3).
Significant interoperability concerns
The hard part is out of band. The IdP has to know the client's client_id at the other server (Section 5.1), the sub that server expects (Section 5.2), and which resources it governs. That last one is "out of scope for this specification" (Section 4.3.3). Every IdP and SaaS vendor pair still needs setup the spec doesn't describe.
Five pages of normative rules sit in Security Considerations. All of the DPoP and mTLS processing is in Section 9.11. An implementer who reads Section 4 and skims Section 9 will miss MUSTs.
The two binding methods behave differently. A DPoP-bound ID-JAG uses grant_type=jwt-dpop. An mTLS-bound one uses jwt-bearer (Section 9.11.2.2). The metadata section only requires jwt-bearer in grant_types_supported (Section 7.2), so a client can't tell ahead of time whether a server accepts bound grants. It finds out when the request fails.
Step-up is not specified. Section 9.3 says an insufficient_user_authentication error "may be returned", in lowercase, with a max_age field in a token endpoint error body. Then comes the TBD.
invalid_request now means two things at the IdP. It covers an expired subject token and a malformed request. The client is told to guess from whether the request "previously succeeded" (Section 4.3.4.3).
ID-JAG reuse is the server's choice. A client MAY resubmit the same ID-JAG. The server may also accept each one only once (Section 4.4.4). The client learns which by getting invalid_grant.
Mechanisms that could be removed
- SAML. The
saml-nameidformat (Section 3.2) and the SAML-to-refresh-token exchange (Section 4.5) take about six pages. They could be a separate document. actor_tokenandact. Both are listed, and the draft "does not define normative processing requirements" for either (Section 4.3).- The protected discovery mechanism in Section 9.7. It carries two MUSTs and has no defined interface.
- mTLS binding, unless someone plans to deploy it.
Writing Quality
The writing is standard IETF style and mostly clear. Requirements usually come with a reason. The Section 9.5 attack walkthrough is a good example.
It is too long. Eighty pages is a lot for two token requests, and twelve of those pages are use cases. The length has started to cause inconsistencies:
- Section 5.1 says the client "still needs to authenticate" at the Resource Authorization Server. Section 4.4 now makes that optional.
- The DPoP example in Section 9.11.1.1 has no client authentication. Section 4.3.3 requires it.
- The numbered list in Section 9.11.1.1 restarts at 1 after the JSON example.
actis an OPTIONAL claim with no processing rules.
Authors
Aaron Parecki (Okta). 2 published RFCs: RFC 9728 (Protected Resource Metadata) and RFC 10017 (OAuth 2.0 for Browser-Based Applications). 18 active Internet-Drafts, including OAuth 2.1, CIMD and ID-JAG. An established author doing a lot of active writing. He is the sole author of the jwt-dpop draft this one depends on.
Karl McGuinness (Independent). Datatracker shows no published RFCs and 25 active Internet-Drafts. Not a new author, and writing at a very high rate.
Brian Campbell (Ping Identity). Datatracker shows 14 published RFCs and 9 active Internet-Drafts. The list includes RFC 7523, Token Exchange, mTLS, and DPoP, which are the specs this draft profiles.
Questions for the Working Group
So what should the working group decide at IETF 127?
- Is the feature set frozen? -05 added mTLS binding and unauthenticated clients. Last call can't start while the scope keeps growing.
- What happens to the
jwt-dpopdependency? Adopt that draft, or pull the grant type into this one. - Should SAML support move to its own document?
- Resolve the Section 9.3 TBD. Specify step-up or cut it.
- Should the sender-constraining rules move out of Security Considerations into Section 4?
- Can this publish with a normative reference to an OpenID draft for
tenantandaud_sub?