The Scope Creep Automated reviews of new AI-related working group submissions

Review · August 21, 2026

WIMSE Applicability for AI Agents

draft-ni-wimse-ai-agent-identity -02View on Datatracker →

Problem Statement

AI agents operate autonomously across multiple systems with minimal human oversight, but existing security models treat them as extensions of user or device identities rather than as first-class principals. This spec argues that agents need their own independent, cryptographically verifiable identities — and proposes how the WIMSE architecture can be applied to provide them, including a mechanism ("dual-identity credentials") that binds an agent's identity to the human or organizational owner it acts on behalf of.

How the Spec Fits In

This is primarily an applicability statement — it maps the existing WIMSE architecture onto AI agent scenarios rather than defining net-new protocol mechanisms. It depends on or references:

  • WIMSE core (the workload identity architecture this whole effort builds on)
  • RATS / EAT (for attestation evidence during credential bootstrapping)
  • BRSKI (RFC 8995) and EST (RFC 7030) (for initial trust establishment)
  • draft-messous-eat-ai-01 (an EAT profile for AI agent claims)

The spec introduces one novel concept: the dual-identity credential, which cryptographically binds an agent identity to an owner identity. It proposes three distinct flows for issuing these credentials (agent-mediated, owner-mediated, server-mediated) and maps the WIMSE "identity proxy" role onto an AI-agent-specific component that mediates credential requests.

It explicitly positions itself as not addressing access permission inheritance (distinguished from CHEQ, which governs action-level authorization) — this spec is about identity provisioning only.

High-Level Summary

The spec introduces an "Identity Proxy" component that sits between AI agents and an Identity Server within a trust domain. Agents don't request credentials directly; they go through this proxy, which adds hardware/OS attestation evidence before forwarding to the server.

The core new idea is the dual-identity credential: a single credential containing cryptographic material for both the agent and its owner (person or organization), binding accountability to both parties. Three issuance models are proposed for how the owner's signature gets incorporated:

  1. Agent-Mediated: The owner signs offline/locally; the agent assembles a composite request. Good for local deployments with hardware keys (TPM), but risky if signing keys are unprotected.
  2. Owner-Mediated (Gateway Mode): The owner sits in the request path as a supervisory gateway, inspecting before forwarding. Good for enterprise governance but creates a single point of failure.
  3. Server-Mediated (Challenge-Response): The server sends an out-of-band challenge to the owner independently. Good when the owner isn't on the request path, but requires careful replay protection.

For bootstrapping initial trust (before any credentials exist), the spec references BRSKI as one viable path, with a note that simpler EST may suffice in pre-configured environments — but leaves the choice open as an explicit unresolved question.

Interoperability Assessment

Significant Interoperability Concerns

The spec presents three issuance models without normative guidance on which to use or when. Implementations will make incompatible choices in practice, undermining the goal of interoperable dual-identity credential issuance. If all three models are meant to be supported, the spec needs to define a negotiation or selection mechanism; if one is preferred, it should be called out explicitly.

The format of the dual-identity credential itself is never defined. Readers learn what it conceptually contains (agent identifier, owner identifier, public keys for both, binding proof) but there is no wire format, no JWT/CBOR mapping, and no reference to an existing credential format being extended. Without this, nothing is actually interoperable.

The Security Considerations section is marked "TODO." For a spec whose entire motivation is improving security, this is a critical gap that blocks any serious evaluation.

The bootstrapping question is posed as an open issue by the authors themselves ("what conditions determine which bootstrap method to use?") — meaning the spec knowingly leaves a core design decision unresolved.

Mechanisms That Could Be Removed

The comparison with CHEQ (Section 4.3) reads as background context, not a normative mechanism. It is useful framing but could be moved to an appendix or cut entirely without affecting the spec's interoperability surface.

The campus network (802.1X / EAP-TLS) use case in Section 4.1 is illustrative but introduces a lot of networking context unrelated to the WIMSE/agent credential protocol itself. It could be trimmed significantly or dropped in favor of a simpler motivating example.

Writing Quality

The document reads more like a position paper or requirements analysis than an IETF protocol specification. It contains no RFC 2119/8174 normative language (no MUST, SHOULD, MAY keywords), which means nothing in the document is actually normatively required — it's all descriptive. For a spec intended to produce interoperable implementations, this is a foundational gap.

Grammatically, the document is readable but has some awkward phrasings and constructions consistent with non-native English writing ("Dual-identity credentials can manifest Human-in-the-Loop oversight" is representative). This isn't a showstopper but would need a writing pass before publication.

The spec is moderately well-structured, with clear section headers and a reasonable flow. However, the risk/mitigation discussions embedded in each issuance model subsection feel more like design notes than specification content, and the open question at the end of the bootstrapping section signals the document isn't yet ready for standardization.

Author Background

Yuan Ni (Huawei) has no published RFCs. All five of their active drafts are in the AI agent security/identity space, including this one, suggesting they are a new and highly focused contributor. The concentration of drafts in one area may reflect a coordinated push from Huawei on AI agent identity rather than a long IETF track record.

Chunchi Peter Liu (Huawei) also has no published RFCs but is significantly more active, with 15 active internet drafts spanning OAuth, AI agent security, IoT/MUD, SATP, ANIMA, ACME, and RATS. They serve as Secretary of the SATP working group, indicating meaningful organizational IETF involvement. The breadth of active drafts is notable — 15 concurrent drafts is a very high volume, which may affect the depth and polish of individual documents.

Both authors are relatively new to the IETF publication track (no RFCs), and both are affiliated with Huawei, which is a factor to note in evaluating the independence of the use cases described.