Review · November 10, 2026
Independent Implementation and Formal Analysis of Attenuating Delegation Chains for AI Agents Summary
The problem
Two individual drafts propose offline-verifiable delegation chains for AI agents. Each token in the chain can only narrow the authority of the one before it. Nobody outside the authors had checked whether the text of either draft is enough to build a verifier.
This draft is that check. An unaffiliated implementer wrote a verifier from each draft's text, built a formal model of both chain structures, and wrote down every place the text left a decision open.
It specifies nothing. There is no protocol, no new claims and no IANA actions.
How it fits in
This is an implementation report, modeled on the "Implementation Status" section from RFC 7942. It isn't a candidate for adoption and doesn't ask to be one.
The two drafts it covers are:
- Attenuating Authorization Tokens (AAT), revisions -01 and -02, submitted to OAuth
- Agent Delegation Chain (asor-02), submitted to WIMSE
Both are individual drafts. Neither is adopted. Half of this report is about a WIMSE draft, so it's only partly an OAuth document.
It leans on the same RFCs the two drafts do: Rich Authorization Requests, DPoP, the JWT access token profile, and a handful of JOSE RFCs. It adds no roles or relationships of its own.
So why read it? It's the best available evidence on how mature these two designs are, and it shows where they actually differ.
What it actually contains
A verifier. A minimal TypeScript implementation with three profiles: AAT-01, AAT-02 and asor-02. It only verifies chains. There is no issuing service, no CBOR and no revocation.
A formal model. A ProVerif model of the chain structure for AAT-01 and asor-02. The attacker controls the network and can compromise any holder key, but never the trust anchor.
Findings on AAT. 25 observations on -01. The -02 text resolves 19 of them, partly resolves one, makes one moot and leaves four open. The -02 verifier profile was written from the text and then run against the AAT author's 229 test vectors. All 229 agreed on the first run, including the exact failing step on all 175 denials.
Ten open points in AAT-02. Places where two conforming verifiers can give different answers. Examples: duplicate members in the JOSE header, whether typ comparison is case-sensitive, whether aat_aud can be an array, and two trust anchors that share one key.
Findings on asor-02. Seven observations, and one of them is serious.
The finding that matters
The two drafts build nearly the same chain. They differ on who signs each link below the root.
In AAT, a child token must be signed by the key its parent names in cnf. Only the holder of a token can extend the chain below it.
In asor-02, the verification algorithm never says which key verifies a link below the root. A verifier can only check each link against its configured set of trusted keys.
This is a real attack vector in a multi-signer deployment. An attacker holding any one configured key can sign a link below someone else's token. It puts its own key in the forged leaf's cnf, presents its own DPoP proof, and the request is accepted. DPoP binds the presenter of the leaf and does nothing about who signed the links above it.
The report frames this as a different trust model, not a defect. The practical result is the same either way. An asor-02 chain is as secure as the weakest key its verifier accepts. The asor-02 author's implementations and all 35 test vectors use a single signing key, so the question never came up.
The model also isolates the cause. Change one rule so that each link verifies under its parent's cnf, and asor-02 gets the same results as AAT. That change turns asor-02 into AAT's design.
Two smaller asor-02 gaps are worth knowing:
- A child's
del_max_depthis checked for monotonicity and never enforced as a limit. A token that says "max depth 1" can still have grandchildren. - The chain length bound is off by one depending on whether
nis read as an index or a count. No test vector reaches the boundary.
What it confirms from list review
Neil Madden's review of AAT-01 on the OAuth list raised two points. The report tests both in the model.
The first point holds. A parent token stays valid after a child is derived from it. This is how both designs work. AAT-02 now says so in its security considerations.
The second point doesn't hold for AAT. An attacker who observes an intermediate token can't extend it, present it, or move it under another parent. It does hold for asor-02, for the reason above.
Interoperability and necessity
There is nothing here to interoperate with, so the usual questions apply to the two drafts under study.
The AAT result is strong but narrow. Matching 229 vectors from the text alone shows the verification algorithm in -02 is specified tightly. The implementer says plainly that the verifier has never exchanged a token with another running implementation. It also skips issuance, derivation, extension constraint types and every algorithm except Ed25519.
The asor-02 result is weaker. The 35 vectors all use one HS256 key and a placeholder cnf, so signature key selection and holder binding are never exercised.
Does this need to be an Internet-Draft? Not really. The findings could have been a mailing list post and a GitHub repository. Publishing it as a draft does give the findings stable labels and a citable revision, and the AAT-02 summary on this site already cites it.
It shouldn't progress toward an RFC. It reports on specific revisions of two unadopted drafts and will be stale when either one revs.
Things that could be trimmed
- The BCP 14 boilerplate. The document has no normative requirements.
- The full ProVerif table. 37 rows, most of them expected outcomes of variant models. The four-row summary table later in the document carries the result.
- Findings that closed without changing anything. B9, B10 and B16 note redundant checks and non-issues. They're useful to the AAT author and noise for everyone else.
Writing quality
The writing is precise and unusually careful about its own limits. There is a section titled "What the Model Does Not Say". Every finding ends with what the implementation chose to do. The report states that a "true" from ProVerif is a proof about a bounded model, not about the draft text.
It's also dense. Findings are written in a compressed shorthand of step numbers and vector IDs (3k, 4p2, J.21.6). You need both drafts open to follow most of it. At 31 pages it's a reference document for the two authors, not something to read straight through.
One formatting bug: finding A3 breaks mid-sentence into a stray numbered list.
The tone is neutral throughout. It doesn't pick a winner between the two drafts, even though its own results favor AAT's signing rule.
Background of the author
Azat Davliatshin is the sole author and lists no affiliation. The draft says the author is not an author of either draft under study.
I couldn't reach Datatracker during this review, so RFC and active draft counts are unverified. This is the only document by this author in the OAuth, WIMSE and related collections tracked here, which points to a new IETF participant.
So what should the working group do with it?
Nothing procedurally. Use it as input.
If you're evaluating AAT, the list of ten underdetermined points is a ready-made review of -02. The author should close them in -03.
If you're evaluating the WIMSE Agent Delegation Chain, the signer question needs an answer before anyone deploys it with more than one signing key.
If you're comparing the two, the report reduces the difference to one rule: whether the holder of a parent token must sign its child. That's the design decision the working groups need to make, and it applies to every other delegation chain draft in this space too.