Review · August 21, 2026
Carrying Remote Attestation Evidence in Workload Identity Tokens (WIT)
Problem Statement
WIMSE defines Workload Identity Tokens (WITs) for conveying workload identity in service-to-service communication, but does not address how to prove that the workload is running in a trusted hardware environment (a Trusted Execution Environment, or TEE). This spec defines how to embed TEE attestation measurements directly in a WIT, enabling a receiving service to quickly verify not just who a workload claims to be, but where it is running and whether that environment is trustworthy.
How the Spec Fits In
This is a targeted extension to the WIMSE s2s protocol (draft-ietf-wimse-s2s-protocol), adding four new JWT claims to an existing token structure. It does not create new OAuth roles or flows. It depends on:
- The WIMSE s2s protocol for the overall WIT definition and DPoP binding
- RFC 9449 (DPoP) for token-to-request binding
- RFC 9334 (RATS Architecture) for the attestation framework and terminology
- RFC 7519 (JWT) and RFC 8725 (JWT BCP) for the token format
The spec sits at the intersection of WIMSE and the RATS working group, applying existing attestation architecture concepts to the WIMSE token format.
High-Level Summary
The spec introduces four new JWT claims that can be included in a WIT:
attested_environment: A boolean flag indicating the workload runs in a TEE capable of producing attestation evidencetee_type: Identifies the specific TEE hardware (e.g., Intel TDX, AMD SEV-SNP, Intel SGX, ARM CCA)measurements: Embedded summary hash values from the TEE's measurement registers, enabling lightweight "fast-path" verificationevidence_ref: A URI pointing to the full RATS evidence for "deep-path" verification when higher assurance is needed
The two-tier verification model is the central design choice: a service receiving a WIT can quickly check the embedded measurement hashes against a local policy without any external calls (fast path), or optionally fetch full cryptographic evidence from the referenced URI when the operation demands it (deep path). The WIT is DPoP-bound as defined in WIMSE, which ties the attestation claims to a specific HTTP request and prevents replay.
The trust model places significant responsibility on the WIT issuer: the issuer is assumed to have already validated the TEE attestation evidence before embedding measurements in the token. The receiving service trusts the issuer's assertion rather than independently verifying the hardware attestation chain.
Interoperability Concerns
Incomplete specifications for three of four TEE types. The spec registers four TEE types in IANA (Intel TDX, AMD SEV-SNP, Intel SGX, ARM CCA) but only fully defines the measurement format for Intel TDX. The AMD SEV-SNP, Intel SGX, and ARM CCA measurement formats are all marked TBD. This means interoperable implementations can only be built for Intel TDX today, while the other three registrations are placeholders. Publishing with this many TBDs in a -00 is understandable, but the spec should not advance until these are filled in.
Deep-path protocol is underspecified. The evidence_ref claim points to a URI where full RATS evidence can be fetched, but the spec does not define the retrieval protocol beyond "HTTPS with TLS 1.2+." There is no specification of the request format, authentication mechanism, response format, or error handling. Interoperability of the deep-path requires at least a minimal protocol profile for this endpoint.
Key binding to TEE is deferred. A critical security property — proving that the DPoP signing key was generated within the attested TEE — is explicitly deferred to future versions. Without this, a compromised workload outside the TEE could potentially present stolen attestation measurements alongside a DPoP key it generated in untrusted software.
Mechanisms that could be removed or deferred. The summary sub-field within measurements (a hash-of-hashes) seems redundant given that individual register values are already present. The spec could remove this or make it purely optional without impacting interoperability. The cross-domain workflow section (Section 5) largely restates WIMSE s2s and DPoP behavior and could be condensed significantly.
Writing Quality
The writing is reasonably clear and structured for a -00 draft, and uses standard IETF RFC 2119 language appropriately. The security considerations section is notably thorough for a first draft, with a reasonable threat model and property table.
However, the three TBD measurement format sections substantially undermine the spec's utility. Registering TEE types in IANA without defining their measurement formats creates incomplete registrations. A narrower initial scope — perhaps just Intel TDX, which is fully specified — would make for a tighter and more reviewable document. The spec would benefit from clarifying the trust model for the WIT issuer's attestation validation, which is load-bearing but described only at a high level in the security considerations.
Background of the Authors
Dapeng Liu ([email protected]) — 3 published RFCs, primarily in the distributed mobility management area from 2014–2015. He has 10 active internet-drafts and an extensive history of IETF participation (81+ expired drafts) spanning mobile networking, IoT, SDN, and now OAuth/WIMSE and AI agent authorization. He is an experienced IETF participant who is currently very active.
Hongru Zhu ([email protected]) — 0 published RFCs, 6 active internet-drafts. Her recent work is focused on OAuth, AI agent authorization, and WIMSE — closely related to this draft. She is a relatively new IETF author but appears active in the WIMSE and OAuth space.
Jian Jin ([email protected]) — 0 published RFCs, 1 active internet-draft (this one). Prior expired drafts relate to DNS extensions. This appears to be his first significant WIMSE/security contribution and he is a new author in this space.