Review · November 16, 2026
AI Identity Management System (AIMS)
The Problem
People keep inventing new authentication and authorization schemes for AI agents. Most of them rebuild things that already exist.
This draft argues that an agent is a workload (Section 4). It then shows how WIMSE, SPIFFE, OAuth and OpenID Shared Signals already cover agent identity, credentials, authentication, authorization and monitoring. It defines no new protocol.
How It Fits In
This is an adopted WIMSE working group draft, not a new submission. It was adopted from draft-klrc-aiagent-auth-03 and is now at -01. It has a 30 minute slot at IETF 127.
It adds no mechanisms, no claims and no registries. Section 16 has no IANA actions. The one new thing is a term: "AIMS", a conceptual model of eight components stacked from identifier up to compliance (Section 5).
The dependency list is long. The draft cites 14 RFCs and 11 IETF drafts, plus SPIFFE and OpenID Foundation documents. All of them sit under "Normative References". The only informative reference is a blog post.
Ten of the eleven drafts are adopted work in WIMSE or OAuth. One is an individual draft, the transaction token chaining profile.
So what stage is it at? Early. The content is a survey with requirements mixed in, and the working group has not yet settled which of those two things it is.
Summary
The agent gets exactly one WIMSE identifier, which may be a SPIFFE ID (Section 6). It gets short-lived credentials bound to that identifier: a WIT, an X.509 certificate, or a JWT-SVID (Section 7). Static API keys are called an antipattern.
Credentials are issued at runtime after a posture check of the agent and its environment (Section 8). The LLM MUST NOT have access to the agent's credentials.
The agent authenticates with mutual TLS, or at the application layer with WIMSE Proof Tokens or HTTP Message Signatures (Section 9).
For authorization the agent is an OAuth client (Section 10). The agent's identity goes in client_id and the user goes in sub (Section 10.3). The draft maps scenarios to existing grants:
- A user delegates: authorization code grant, with the agent authenticating using its workload credential instead of a client secret (Section 10.5.1).
- The agent acts for itself: client credentials or a JWT authorization grant (Section 10.5.2).
- Inside a service: Transaction Tokens (Section 10.6).
- Across domains: Identity Chaining or the Identity Assertion JWT Authorization Grant (Section 10.7).
- A human must approve: CIBA (Section 10.8).
Section 11 covers revocation signals over the Shared Signals Framework and requires audit logs.
What changed in -01
The text grew from about 1,680 lines to about 1,850. The changes answer issues filed after adoption.
- New Section 9.3 on long-lived sessions and streams. A session is bounded by the credential that opened it and is re-validated when that credential expires.
- New Section 10.4 on policy systems. A common policy language is not required across trust domains. This section also adds sender-constrained tokens using mutual TLS or DPoP.
- Section 8 adds deprovisioning, and says posture checked at issuance may not hold at request time.
- Section 10.6 says a token obtained through exchange can never exceed the scope of the token presented.
- Section 11 adds the OpenID WISE profile.
Interoperability
Significant interoperability concerns
It does not pick. Section 7 lists five credential formats and says the choice "depends on the deployment's trust model". Section 9 offers three authentication mechanisms. Section 10.3 does not require a token format. Two deployments can follow every line of this document and have no credential, authentication method, or token format in common. A survey can live with that. A profile cannot.
The document has not decided what it is. The abstract says it "proposes best practices". Section 2 says it "profiles existing standards" and "defines best current practices". The header says Informational. It uses BCP 14 keywords more than 40 times.
Many of the MUSTs cannot be tested. Deployments "MUST maintain sufficient monitoring", audit records "MUST be tamper-evident", revocation must be enforced "without undue delay" (Section 11). None of these change what goes on the wire.
The human-in-the-loop requirement has no mechanism behind it. Section 10.8 says a local confirmation prompt MUST be bound to a grant from the authorization server, and suggests CIBA. The note in the same section says CIBA "doesn't map well" to confirmation mid-execution. This is the clearest gap the draft has found, and it is recorded as a note.
client_id as agent identity is thin. Section 10.3 puts the agent in client_id. Section 10.7 sends agents across domains with token exchange. The draft never mentions the act claim, so it does not say how a chain of agents shows up in a token.
Mechanisms that could be removed
- Section 13 on compliance. It is one paragraph that declares itself out of scope.
- The A2A, ACP and AP2 references. They are listed as normative and never cited in the text.
Writing Quality
The structure is clear and the survey is useful. A reader new to this space gets an accurate map in 25 pages.
The text needs an editing pass:
- Section 8 says provisioning "typically includes two phases" and then lists three.
- Section 9.2.2 explains what
@request-targetcovers. The current HTTP Signatures draft (-08) replaced@request-targetwith@pathand@query. - Section 10.4 has a sentence fragment starting "The authenticated identity of the requester".
- The transaction token chaining profile is cited under an old draft name. Client ID Metadata Document is cited at -02 and the current version is -03.
- Typos: "requriements", "langugage", "Oauth protocol famility".
Authors
From their Datatracker profiles:
- Pieter Kasselman (Defakto Security). 1 RFC, RFC 10027. 7 active drafts. WIMSE chair.
- Jeff Lombardo (AWS). No RFCs. This is his only active draft.
- Yaroslav Rosomakho (Zscaler). 1 RFC, RFC 9850. 28 active drafts across WIMSE, TLS, MASQUE and OAuth. IAB member.
- Brian Campbell (Ping Identity). 14 RFCs, including Token Exchange, mutual TLS and DPoP. 9 active drafts.
- Nick Steele (OpenAI). No RFCs. 2 active drafts.
- Aaron Parecki (Okta). 2 RFCs, RFC 9728 and RFC 10017. 18 active drafts.
Questions for the Working Group
So what should the 30 minutes be spent on?
- Is this an Informational framework or a BCP? If Informational, should the normative keywords stay?
- Should it pick a minimum set, one credential format and one authentication method, or stay a menu?
- Should the gaps get their own section? Mid-execution user confirmation is one. Agent delegation chains are another. The introduction promises to "identify gaps", and right now they are scattered in notes.
- Which references are really normative? Publication would otherwise wait on eleven drafts.