Review · November 10, 2026
Delegation Chaining for the Identity Assertion JWT Authorization Grant
ID-JAG covers one hop: a client trades an identity assertion for a grant it redeems at another authorization server. It accepts an actor_token but says nothing about what to do with it. Section 4.3 leaves that to "future profiles."
This draft is one of those profiles. It lets an agent holding an ID-JAG trade it for a new, narrower ID-JAG bound to the next agent in the chain. The final access token carries the whole chain (User, Agent A, Agent B, Agent C) in nested act claims.
How it fits in the working group
It's a profile of ID-JAG, which is a profile of Identity Chaining. Both are adopted work, and ID-JAG explicitly invites this kind of extension. That's within the charter.
It defines no new parameters, claims, endpoints or grant types. There are no IANA actions. Everything is assembled from Token Exchange, RFC 7523, Resource Indicators and Rich Authorization Requests.
Normative dependencies are seven RFCs, the two working group drafts above, and three SPIFFE documents pinned to a GitHub commit. SPIFFE client authentication is informative but recommended. WIMSE is not mentioned.
It does stretch ID-JAG in three ways:
- An ID-JAG becomes a valid
subject_token. ID-JAG only lists ID tokens, SAML assertions and refresh tokens. audiencecan name the same AS that issues the ID-JAG. ID-JAG defines it as the Resource AS in another trust domain. The AS ends up issuing grants to itself.actor_tokennames the next hop instead of the party making the request.
So what new relationships does this create? Two. A downstream agent hands a credential to its upstream peer. In the cross-domain case, the upstream AS needs the SPIFFE trust bundle of the downstream domain. That is a different trust path from the AS-to-AS trust ID-JAG already assumes.
It also lands in a crowded space. Actor Chain, Transaction Tokens for Agents and the same author's earlier Delegation Chain draft all record multi-hop agent lineage. This draft cites none of them.
What the mechanism actually does
- Agent A gets an initial ID-JAG for the user, the normal ID-JAG way.
- Agent A wants to delegate to Agent B. It asks B for B's JWT-SVID, minted with the AS as its audience.
- Agent A authenticates to the AS and sends a Token Exchange request.
subject_tokenis A's ID-JAG.actor_tokenis B's JWT-SVID.scopeandauthorization_detailsare the same or narrower. - The AS checks that A is the client named in the ID-JAG's
may_act, checks that nothing got wider, applies its delegation policy and enforces a depth limit. - The AS issues a new ID-JAG.
subis still the user.actrecords A.may_actandclient_idname B. - A passes the ID-JAG to B. B repeats the process, or redeems the ID-JAG for an access token with the JWT bearer grant.
Every hop is one call to an AS, and the AS signs every ID-JAG. No agent mints its own credentials.
Section 5 then adds an optional variant that does the opposite. Agents self-issue "Delegation Assertions" signed with their X.509-SVID key, and only the last agent talks to the AS.
Significant interoperability concerns
Agent B hands Agent A its own client credential. The JWT-SVID that B gives to A has the AS as its audience. That is exactly what B uses to authenticate to that AS under SPIFFE client authentication. The draft claims "the delegating entity itself cannot redeem the ID-JAG it requested for its target." A holds both the ID-JAG and B's credential, so A can do exactly that. The defense is a MUST to reject a reused jti. JWT-SVIDs aren't required to have one. The SPIFFE client authentication draft says outright that a party who obtains a JWT-SVID can replay it for client authentication until it expires.
The actor_token isn't needed. The draft says the delegation policy check, "and not the actor_token, is what authorizes a hop." The AS only needs B's identifier to fill in may_act. A plain parameter naming the target would remove the credential handoff, the extra round trip to B, and the cross-domain trust bundle requirement.
act is one hop behind. RFC 8693 says the outermost act is the current actor. In the ID-JAG that B holds, the outermost act is A. The draft acknowledges this differs from the RFC 8693 examples and calls it AS discretion. Any component that reads act the standard way gets the wrong answer until the final access token.
Only SPIFFE deployments get a complete spec. Other client authentication methods "MAY be used." may_act.sub and client_id are always SPIFFE IDs taken from the JWT-SVID. How the AS matches those against a client that authenticated with private_key_jwt and a regular client_id is not defined.
Single-use ID-JAGs prevent fan-out. The AS MUST reject a subject_token whose jti it has already seen. Agent A can delegate to exactly one agent, once. The appendix has A request reports:read mail:send, which suggests two downstream tasks. The second exchange fails.
60-second lifetimes don't match agent workloads. Each issued ID-JAG SHOULD NOT live longer than 60 seconds. If Agent B works for two minutes before deciding to call Agent C, its ID-JAG is dead. B has no way to refresh it. It has to go back to A.
RAR narrowing is left to each type. Whether one authorization_details object is narrower than another "is defined by that authorization_details type." No existing type defines that. The AS MUST reject when it can't decide, so RAR works only with types both sides have privately agreed on.
No discovery and no error codes. A client can't tell from AS metadata whether the AS accepts an ID-JAG as a subject_token. Policy denial, depth exceeded and may_act mismatch all get an unspecified rejection. The agent can't tell "try a narrower scope" from "this chain is too deep."
Depth counting conflicts with the privacy advice. The AS counts nested act claims to enforce the depth limit. The privacy section says an AS SHOULD truncate the act chain when topology is sensitive. A truncated chain resets the count for the next AS.
Mechanisms that could be removed
All of Section 5. The self-issued variant undoes the property the rest of the draft is built on. The draft lists the problems itself: the AS "cannot tell from the assertion alone whether its chain of issuers truly roots at a grant the AS issued," attenuation is "not independently verifiable," and "any holder can attempt redemption."
In practice, any workload with a valid X.509-SVID can sign an assertion naming any user as sub. The only check is a SHOULD to "corroborate the root by means available."
The section is also underspecified. The certificate goes in x5c "for example." The token type is the generic jwt. agent_context is listed as a claim, and two pages later the draft says it "does not define it." Its example contains delegation_depth, right after Section 4.4 explains why no depth claim is needed.
Section 5.1 says the variant SHOULD NOT be used when the AS is reachable. Section 6.2 then recommends it for deployments that don't want to share a JWT-SVID. Removing the actor_token fixes that problem without a second protocol.
The SPIFFE requirement. The mechanism is a chain of may_act and act claims enforced by client authentication. Nothing about that needs a specific workload identity format. Requiring only an authenticated client and a client identifier for the target would cover SPIFFE, WIMSE and ordinary OAuth clients.
What's left is a small and useful profile: ID-JAG as a subject_token, may_act binding to the next client, nested act, and attenuation rules.
Writing quality
The writing is good. Sentences are clean, BCP 14 usage is careful, and the draft cites specific sections of RFC 8693 and ID-JAG whenever it departs from them. The security considerations are honest about what each mechanism does not protect against.
It's 31 pages for a profile with no new protocol elements, which is long. The same points are restated many times. "Exactly one request to an AS" and "the act claim is built from the authenticated client identity" each appear four or five times. Step 4 and step 6 of the processing rules interleave delegation and redemption cases in single paragraphs and are hard to follow. Splitting them into two short lists would help.
The end-to-end appendix is useful and matches the normative text.
Background of the Authors
Datatracker was unreachable for this review. The numbers below come from our earlier 2026 reviews.
- Dapeng Liu (Alibaba): 3 RFCs from 2014 and 2015 in distributed mobility management, around 10 active drafts and 80+ expired ones. An experienced author who is writing a lot right now across OAuth, WIMSE and agent authorization. This draft uses a new Alibaba email address, so it may show up under a separate Datatracker profile.
Liu is the sole author. Aaron Parecki (Okta), a co-author of ID-JAG and Identity Chaining, is listed as a contributor. Hongru Zhu, Hui Xue and Suresh Krishnan are acknowledged.