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

Review · August 20, 2026

Condition-Bound Keys for Mutual-TLS Client Authentication and DPoP

draft-winmagic-oauth-condition-bound-keys -00View on Datatracker →

Problem Statement

OAuth's sender-constraining mechanisms (mTLS, DPoP) prove only that a key was possessed at the time of a request — they say nothing about whether the user is still present or the platform is still in a compliant state. This spec proposes "condition-bound keys": hardware-rooted, non-extractable keys that can only be used (i.e., can only sign) while a set of locally-evaluated conditions are satisfied. When a condition fails — user walks away, platform falls out of policy — no new proof can be generated. The key effectively disappears without any revocation signal.


How the Spec Fits In

This is a conceptual framework/profile spec that layers on top of RFC 8705 (mTLS) and RFC 9449 (DPoP) without modifying either protocol. It does not define new wire formats, new parameters, or new endpoints. What it does is define a semantic layer: if the key behind an existing mTLS or DPoP proof has certain hardware-enforced properties, then a successful proof implies more than it did before.

The spec also touches on Device Bound Session Credentials (DBSC), WebAuthn/FIDO2, and Continuous Access Evaluation Profile (CAEP/SSF) — making it broader in scope than a typical focused OAuth draft. It references RFC 9334 (RATS attestation) as the recommended mechanism for conveying evidence that a key has these properties.

The authors also have a companion draft (draft-winmagic-wimse-condition-bounded-credentials) covering the same concept for workload identity in WIMSE. This OAuth draft handles the human principal case.

Dependencies include: RFC 8705, RFC 9449, RFC 9334, RFC 8693 (token exchange), RFC 7250 (raw public keys), and several transport RFCs (9113, 9001, 9846).


High-Level Summary

Today, when a client presents an OAuth token bound to a certificate or DPoP key, the resource server can verify that the caller controlled that key at the moment of the request. But it cannot verify that the user who authorized the token is still at the keyboard, or that the device is still meeting corporate policy. Token lifetimes serve as a rough proxy for freshness of these conditions.

This spec introduces "condition-bound keys" — cryptographic private keys generated inside hardware (TPM, Secure Enclave, TEE) where:

  1. The key cannot be extracted in usable form.
  2. The key can only be used (i.e., sign) when locally evaluated policy conditions are met (e.g., user is biometrically verified as present, platform is compliant).
  3. The enforcement is hardware-gated, not software-asserted.
  4. The authorization server can obtain evidence (via attestation) that the key actually has these properties.

When this kind of key is used as the binding key in mTLS or DPoP, the existing proof mechanisms inherit a stronger meaning: "this proof exists" implies "the conditions held at the moment it was generated." No new protocol flows are needed at the resource server — it still just checks the thumbprint or verifies the DPoP signature. The difference is what a successful check means.

A key implication: token lifetime can be decoupled from "is the user still there?" Token lifetimes can be set based on how long a permission should last, while user presence is enforced continuously by the key itself.

The spec carefully delineates what can be verified at each party (AS vs. RS), handles edge cases like TLS session resumption and connection reuse, and distinguishes between local conditions (enforced by key absence, no signal needed) and remote conditions (still require CAEP/SSF signals).


Interoperability and Necessity of the Spec

Is enough specified for interoperability? Not yet — and the spec is honest about this. It is a base framework that requires concrete profiles to be written on top of it. The spec repeatedly uses "profiles MUST specify" for critical decisions:

  • Which hardware boundary is acceptable
  • Which conditions gate key use
  • How attestation evidence is conveyed and evaluated
  • How TLS session resumption is handled
  • What the DPoP iat acceptance window is

This is architecturally appropriate for a first-level framework, but it means the spec as written is not directly implementable. Two implementers reading only this document would make many different, incompatible choices.

Are there mechanisms that could be removed? The sections on DBSC, WebAuthn/FIDO2, and the extended "consequence for sessions" discussion (Section 13) are informative and interesting, but they are outside the OAuth working group's scope and could be removed or split out without losing anything normative. Keeping the spec focused on OAuth (mTLS + DPoP) would tighten its scope considerably.

The privacy considerations section is notably thorough and well-reasoned — it is not over-specified, and the guidance on key rotation and attestation evidence scoping is genuinely useful.

What's missing for the spec to be useful? The spec does not define any mechanism for the AS to actually receive or evaluate attestation evidence that a key is condition-bound. It says this should follow RFC 9334, but does not sketch even a minimal profile for how that works in an OAuth client authentication flow. Without this, the spec's central claim — that the AS can establish that a key is condition-bound — has no specified path to implementation.


Writing Quality

The writing quality is unusually good for a -00 draft, and notably better than the typical first submission. The prose is clear, the tables are helpful, and the use of MUST/SHOULD/MAY is mostly disciplined and well-placed. The spec is careful to distinguish informative from normative content.

However, the spec reads more like a position paper or design document than a typical IETF protocol spec. It is rich with explanation and reasoning — which is good — but it spends considerable space on philosophical framing ("login and session become one thing") that, while compelling, is not normative content and adds length without adding interoperability guidance.

Some sections (particularly 12 on the Verified Human Actor and 13 on Consequences for Sessions) are thought-provoking but are entirely informative and could be substantially shortened or moved to an appendix. The spec could be 20–30% shorter without losing any normative content.

The cross-referencing is solid. Security and privacy considerations are genuinely substantive rather than boilerplate.


Background of the Authors

All three authors are from WinMagic, a data security and encryption software company, and all three appear to be new to the IETF:

Thi Nguyen-Huu — 0 published RFCs. 2 active drafts (both WinMagic-authored). No other IETF roles listed. New author.

Sergei Nikitin — 0 published RFCs. 2 active drafts (same two WinMagic drafts). No other IETF roles listed. New author.

John O'Leary — 0 published RFCs. 2 active drafts (same two WinMagic drafts). No other IETF roles listed. New author.

This is an all-new team to the IETF, submitting what is effectively a company research project as a pair of individual drafts. The ideas are technically substantive and the writing is better than average for a first submission, but the authors will benefit from significant community engagement to understand the working group's expectations around scope, the relationship to existing work (particularly WIMSE, attestation, and CAEP), and what it would take to move from a conceptual framework to a concrete, implementable spec.


Summary Assessment

This is an interesting and intellectually serious draft making a valid point: sender-constraining an OAuth token to a hardware-bound, policy-gated key provides stronger guarantees than existing mTLS/DPoP bindings, because it enforces conditions that current mechanisms can only describe at issuance time. The framing is sound and the security/privacy analysis is careful.

The main questions the working group should consider:

  1. Scope: Does the OAuth WG want to own this, or does it belong jointly with WIMSE, RATS, or another working group? The companion WIMSE draft suggests the authors see this as cross-cutting.

  2. Completeness: The spec as written is a framework that requires profiles. Should the authors write at least one concrete profile alongside the framework, or is the framework itself the right submission vehicle?

  3. Relationship to attestation: The connection to RFC 9334 is mentioned but not specified. What does it look like, concretely, for a client to prove to an AS that its key is condition-bound? This is the critical missing piece.

  4. Working group adoption: The spec is currently an individual draft. It would benefit from a working group session presentation before adoption to gauge whether the community sees this as filling a real gap or as complexity in search of a use case.