Review · August 21, 2026
draft-sweeney-wimse-credential-delegation-00 Review
Problem Statement
OAuth 2.0 was designed around human users interacting with browsers. Autonomous AI agents are ephemeral, potentially numerous, and—uniquely—adversarially promptable via prompt injection. This draft addresses how AI agents can be delegated access to protected resources on behalf of human users without ever possessing the underlying credentials, with enforced capability attenuation at each delegation hop, real-time revocation cascades, and a cryptographically verifiable audit chain.
How the Spec Fits In
This is a profile of existing standards rather than a standalone new mechanism. It assembles RFC 8693 (Token Exchange), RFC 9449 (DPoP), RFC 9396 (Rich Authorization Requests), and OpenID Connect CIBA into a coherent framework for AI agent delegation. Despite claiming not to define new token formats or grant types, the IANA Considerations section registers urn:ietf:params:oauth:grant-type:cred-delegation—a contradiction worth flagging.
The spec introduces one significant new architectural role: the Delegation Server (DS), a trusted intermediary that holds credentials in a vault, issues DPoP-bound delegation tokens to agents, and acts as a proxy for every API call the agent makes. This is a meaningful architectural addition beyond the existing OAuth roles.
The spec depends on: RFC 6749, RFC 7523, RFC 8693, RFC 9396, RFC 9449, OIDC CIBA, W3C DID-CORE, and DID-KEY. The W3C DID dependencies are notable in an IETF context and may be controversial. It also references several non-adopted individual drafts (draft-klrc-aiagent-auth, draft-niyikiza-oauth-attenuating-agent-tokens, draft-hartman-credential-broker-4-agents) as related work, indicating this is positioned in a crowded and unsettled space.
High-Level Summary
The spec introduces a three-tier architecture: Subject → Delegation Server → Agent.
A human user (Subject) deposits credentials into the DS's Credential Vault, identified by opaque handles. When an AI agent wants to act on behalf of the user, it authenticates to the DS using an ephemeral key pair (expressed as a DID:key), receives a DPoP-bound Delegation Token (DT) scoped to specific operations and resources via RFC 9396 authorization_details, and then calls the DS's /exercise endpoint rather than calling the resource server directly. The DS validates the DT, retrieves the credential from the vault, makes the actual API call, and returns only the response. The agent never sees the raw credential.
For long-running or asynchronous agent tasks requiring new capabilities, the DS uses CIBA to asynchronously request user consent on the user's registered device before issuing a DT.
When an agent delegates to a sub-agent, the DS signs a receipt linking parent and child DT identifiers and agent DIDs, producing a tamper-evident chain. On revocation, the DS must invalidate the directly revoked DT within 1 second and cascade to all descendant DTs within 5 seconds.
Interoperability and Necessity of the Spec
Significant interoperability gap: the Exercise Proxy.
The most operationally important mechanism—Section 6.2, Proxy Semantics—is substantially underspecified. The spec states the agent's HTTP request specifies an operation and parameters, and "DS maps these to the upstream call," but provides no specification for how this mapping is expressed, negotiated, or structured. Every API the DS must proxy requires the DS to understand that API's request format. Without a standardized way to express "call this upstream endpoint with these parameters using this credential," implementations will make incompatible choices. This is arguably the most critical interoperability surface in the entire spec and it is left open.
The credential wrapping architecture may be over-specified for interoperability.
The Credential Vault, Exercise Proxy, and the strict agent-side network boundary are interesting security properties, but they impose a very specific centralized deployment model on the Delegation Server. Whether a DS stores credentials as opaque blobs vs. encrypted structures, and how it actually invokes upstream APIs on behalf of agents, are implementation decisions that don't need to be mandated for client-DS interoperability. The spec could define the DT format, the /exercise request/response interface, and revocation signals without prescribing the DS internals.
Normative revocation SLAs (1 second / 5 seconds) are likely unimplementable in distributed systems.
These timings are stated as MUST-level requirements. At any non-trivial scale—federation, geo-distribution, clustered deployments—guaranteeing cascade revocation in 5 seconds is architecturally constraining in ways the spec does not acknowledge. Either these should be softened to SHOULD, or the spec needs to explain the deployment assumptions that make them achievable.
Protocol version negotiation (Section 3.1) is premature.
A version negotiation header and HTTP 426 error code in a v00 individual draft adds complexity without adding interoperability value at this stage. This can be deferred.
Mechanisms that could be removed to keep the spec focused:
- Protocol version negotiation (Section 3.1)
- Agent lifecycle management (Section 4.3) — out of scope for a protocol spec
- Consent record retention duration (90 days) — an operational policy, not an interoperability requirement
- The
agent_checksumclaim — noted as "advisory only due to IPR," which raises questions about whether it should be in a normative spec at all
Writing Quality
The writing quality is better than typical for a v00 individual draft. The structure is logical, the problem statement is clear, and Section 1.2's positioning against related work is unusually thorough. The design principles in Section 1.3 read as coherent and internally consistent.
However, there are issues:
The spec claims in the abstract to "not define new token formats, new OAuth grant types, or modifications to existing authorization server behavior," but then the IANA Considerations section registers a new grant type URI. This contradiction should be resolved.
Several sections read more like design documentation than normative specification. For example, Section 3 describes DS internal subsystems (Delegation Engine, Consent Manager, Exercise Proxy, Revocation & Audit) in architectural terms that are useful for understanding but would not constrain interoperable implementations.
The security consideration in Section 9.2 references "a vulnerability disclosed to the OAuth WG on February 26, 2026"—citing a real-world disclosure date in a normative spec is unusual and the reference should either be formalized (an I-D, erratum, or mailing list archive) or removed.
The agent_checksum claim being "advisory only due to IPR" is a significant red flag. IPR concerns affecting a claim's normative weight should be resolved before publication, not footnoted in the claim description.
Background of the Authors
Kieran Sweeney ([email protected])
- RFCs published: 0
- Active Internet-Drafts: 1 (this document)
- New author: Yes, this appears to be a first IETF document submission
This is a new voice in the IETF. The spec is ambitious in scope for a debut individual draft, and the writing quality suggests genuine familiarity with the OAuth and WIMSE specifications. The main gap is the depth of operational experience that typically surfaces concerns like the revocation SLA requirements or the proxy semantics underspecification.