Review · November 20, 2026
Verifiable Attenuated Delegation for AI Agent Chains
The Problem
AI agents hand sub-tasks to other agents. Each hand-off should pass along less authority than the delegating agent holds.
Token Exchange records prior actors in a nested act claim, but that claim is audit history. Nothing lets a resource server check that the token at hop 3 carries no more authority than the token at hop 2. Token Exchange also normally calls the authorization server at every hop.
This draft wants a resource server to verify the whole chain offline, back to a trusted root key, with no call to an authorization server.
How It Fits In
The draft is named for WIMSE, but nearly everything in it is OAuth. It profiles JWT access tokens, carries authority in Rich Authorization Requests, requires DPoP, and uses the Token Status List draft for revocation. It has 16 normative references. The only WIMSE reference used in the mechanism is an optional identifier format for client_id.
It is a new token profile and a new verification algorithm. It does not extend any existing endpoint, and it defines no new endpoint.
It does change a relationship between OAuth roles. Today the authorization server decides what a token can do, and the resource server trusts that decision. Here the resource server receives every token in the chain and recomputes the subset check itself.
The space is crowded. The draft says it wants to converge with draft-niyikiza-oauth-attenuating-agent-tokens, which defines the same three claim names. The two drafts already disagree on del_max_depth. The same depth values can pass in one and fail in the other.
How It Works
A chain is an ordered list of JWT access tokens, DT_0 through DT_n. DT_0 is signed by a trusted root key.
Each token carries:
sub: the accountable principal, identical at every hopclient_id: the agent acting at this hopauthorization_details: exactly one object of typeagent_delegation, holding ascopesarray and aconstraintsarraycnf: the holder's key, proven with DPoPdel_depthanddel_max_depth: position in the chain and the depth limitpar_hash: a SHA-256 hash of the parent token's signing input
The par_hash ties a child to one specific parent token. An attacker can't swap in a broader parent.
Scopes come in three forms: dotted literals (crm.read), trailing wildcards (crm.*), and opaque strings that only match byte for byte. Constraints are typed: max, min, one_of, not_one_of, prefix, rank, plus two cumulative types, max_lifetime and max_subtree.
The resource server runs eight steps. It checks signatures, parent hashes, depth and principal, the subset rule at each hop, expiry, the DPoP proof, revocation status, and then the requested action against the last token.
The draft is candid about what the chain can't do. A per-action max of 100 can be copied to any number of children, so total spend is unbounded. Cumulative budgets need a ledger outside the token. The chain also can't tell you whether the principal still holds the authority the root granted.
Significant Interoperability Concerns
The draft never says who signs a child token. Step 1 says to verify every token's signature and that DT_0 verifies under a root key. It says nothing about which key verifies DT_1 through DT_n. This is the most important question in an offline delegation design.
If the parent's holder signs the child with its cnf key, the draft needs to say so and define what iss contains. If the authorization server signs every token, then a server is in the path at every hop, and the offline claim only applies to the resource server. The reference implementations use a "single-signer trust model", which suggests the second reading.
There is no issuance protocol. The draft defines what a valid child looks like. It doesn't define how an agent gets one.
There is no way to present a chain. An HTTP request carries one access token in the Authorization header. The draft doesn't say where the other tokens go. DPoP's ath binds a proof to one token, and the draft doesn't say which one.
Actions aren't mapped to requests. Step 8 checks "action A" against scopes and constraints. Nothing says how a resource server derives a scope or a max_rows value from an HTTP request. The draft admits the scope vocabulary gap in an open issue. Constraint keys have the same gap and aren't mentioned.
So what does interoperate? The chain validity check does, steps 1 through 5. Two verifiers given the same list of tokens will agree on whether each hop is a subset of its parent. That is a real result, and the test vectors back it up.
The implementation status section shows the limits. Both reference libraries skip holder binding and revocation. The test vectors are signed with HS256, while the draft mandates Ed25519. The libraries still emit EdDSA where the draft requires Ed25519. One independent harness reports 20 passes, 5 failures, and 10 declared gaps against the -02 vectors.
Mechanisms That Could Be Removed
- JCS canonicalization.
par_hashcovers the parent's signing input, which is already a fixed byte string. Requiring JCS on the header and payload adds a rejection rule with no security benefit. Thec14nheader that verifiers "MUST NOT rely on" can go with it. del_depth. It always equals the token's position in the presented chain.- The outcome name table. Error names that never appear on the wire belong in the test suite.
- Section 6.4. It records trust anchor provenance as an open issue, then spends most of a page on another individual draft's audit record format.
- The budget and capacity planning text. Sections 9.4 and 9.5 explain ledger design and authorization server sizing. "A deployment MUST size its authorization server" is not a protocol requirement.
max_lifetimeandmax_subtree. The draft says an offline verifier can't enforce them. They could move to a companion document with the ledger that enforces them.
Writing Quality
The sentences are well formed, and the draft uses IETF conventions correctly. The author is careful to state what the mechanism doesn't prove, which many agent drafts skip.
It is too long. It runs 38 pages for a token format and an eight-step algorithm. Much of the text reads like a log of answers to reviewers. Normative sections credit individuals by name, and the acknowledgments run two pages.
Some text describes the document's history instead of the protocol. "This revision defines no transition mode for chains shaped by the -01" is one example. Four revisions landed in two months, and compatibility with a six-week-old individual draft is already shaping the test vectors.
The length hides the gaps. A reader gets a detailed treatment of re-issuance bursts before noticing that the signer of a child token is never named.
The Author
Rafael Asor (Attenu) is the sole author. He has no RFCs. This is his only Internet-Draft, first posted in August 2026. He maintains both reference implementations, attenu-guard and attenu-guard-ts.
The Datatracker history for the draft also lists an IPR disclosure naming it, posted on 2026-11-03. Anyone considering building on this should read that disclosure first.
Bottom Line
The subset check is carefully specified and testable. It is also only the middle of the problem. Who mints a child token, how a chain reaches the resource server, and what an action means are all left out.
The working groups have at least four overlapping drafts on attenuated agent delegation. This one is better read as a verification algorithm that could feed into a merged effort with draft-niyikiza-oauth-attenuating-agent-tokens than as a standalone protocol.