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

Review · August 20, 2026

Authorization Posture Mechanism (APM) Summary

draft-vicente-oauth-apm -02View on Datatracker →

Problem Statement

Existing OAuth mechanisms (certificate binding, DPoP, step-up authentication) make security decisions at token issuance or at coarse session granularity. APM proposes to re-evaluate three factors — the bound client certificate, the access token, and a device posture signal — on a per-request basis for privileged operations, returning one of four "graduated outcomes" (permit, scope reduction, method restriction, or full denial) rather than a binary allow/deny decision.

How the Spec Fits In

This spec is additive on top of multiple existing OAuth and related RFCs:

  • RFC 8705 (mTLS certificate binding) — APM re-checks the certificate thumbprint on every request, not just at issuance
  • RFC 9449 (DPoP) — APM adds a device posture consistency layer on top of the DPoP key binding
  • RFC 9470 (step-up authentication) — APM replaces the binary "step up or deny" with graduated intermediate outcomes
  • RFC 9396 (Rich Authorization Requests) — APM uses the authorization_details wire format to convey graduated outcomes
  • CAEP / SSF — APM provides a synchronous per-request counterpart to CAEP's asynchronous event push model

It creates a new role, the "APM Enforcement Point," which sits at the resource server and assembles a "Consistency View" on each request. This is a meaningful new architectural component not previously defined by OAuth.

High-Level Summary

When a client presents a request to a resource server, the APM Enforcement Point gathers three inputs: the client certificate thumbprint, the access token's claims, and a device posture signal (an externally issued, integrity-protected assertion about device health). It runs a five-step state machine: verifying the certificate binding, verifying the DPoP key binding (if applicable), comparing the current posture against the posture snapshot recorded at token issuance, and then dispatching one of four outcomes.

Critically, a "downgrade" outcome does not revoke the token — the same token can yield full access on the next request if posture improves. Outcomes are returned to clients using the authorization_details format with a new urn:apm:graduated-outcome:v1 type. The spec maps each outcome class to HTTP status codes (200, 401, 403, 405).

The design explicitly maps to NIST SP 800-207 Zero Trust Architecture components (Policy Engine, Policy Enforcement Point, Policy Administrator), positioning APM as a ZTA implementation pattern.

Interoperability and Necessity

Significant interoperability concerns:

The two most critical functions — ComputePostureDegradation (deciding how degraded the posture is) and DispatchGraduatedOutcome (mapping that to an outcome class) — are explicitly left implementation-defined. The spec states that scoring weights and thresholds are intentionally unspecified. This is a serious gap: two implementations of APM will make entirely different decisions for identical inputs, undermining the core premise of "deterministic" evaluation.

The posture signal format is also not standardized. The spec recommends a baseline JWT shape but does not mandate it. This means there is no interoperable way for a device, an authorization server, and a resource server from different vendors to exchange posture signals.

Mechanisms that could be removed or scoped out:

  • The NIST SP 800-207 ZTA mapping adds length but contributes nothing to interoperability — it is explanatory material better suited to a companion informational document or a blog post.
  • The reason_code field is defined as opaque and explicitly prohibited from encoding useful information. Its value to interoperability is unclear.
  • The implementation status section (describing a private Go prototype) has no place in a standards-track draft.

Core question: Without specifying either the posture signal format or the decision logic, what exactly does this spec standardize that enables interoperability? The wire format for the outcome is specified, but the inputs and the decision are not. A client receiving a scope_reduction outcome has no way to know why or how to remediate, because the spec prohibits the server from conveying that information in the reason_code.

Writing Quality

The writing is organized and uses reasonable section structure. IETF normative language (MUST, SHOULD, etc.) is used, though sometimes inconsistently — some requirements use RFC 2119 keywords and others are written as declarative statements without them.

The spec would benefit from a tighter scope. Several sections (the ZTA mapping, the relationship-to-prior-art table, the implementation status section) read as justification material rather than specification text. These inflate the document without contributing to implementability. The spec also conflates two distinct problems — the per-request consistency check (which is well-motivated) and the graduated outcome dispatch (which is left undefined) — without clearly distinguishing what is and isn't being standardized.

The abstract claims APM defines "deterministic, least-privilege Graduated Outcomes," but the determinism claim is undermined by leaving the core decision logic implementation-defined. This is a material inconsistency that reviewers will flag.

Author Background

Brian Vicente — Sanctum SecOps LLC ([email protected])

  • RFCs published: 0
  • New IETF author: Yes — no published RFCs and no record of expired drafts, indicating very early participation
  • Active Internet-Drafts: 5 drafts currently active, spanning OAuth, ACME, post-quantum cryptography, and PKI

Vicente appears to be a new and prolific draft author exploring several problem spaces simultaneously. The breadth of active drafts (five across very different topic areas) suggests these may be exploratory proposals rather than work backed by existing WG interest or implementation experience.

Overall Assessment

The problem APM is trying to solve — per-request posture-aware enforcement with graduated responses — is genuinely interesting and relevant to enterprise and ZTA deployment scenarios. However, the current draft leaves the two most important pieces (posture signal format and decision logic) entirely implementation-defined, which means it cannot produce interoperable implementations as written.

For this to be viable as an RFC, it would need to either: (a) specify a concrete posture signal schema and a normative decision algorithm, or (b) narrow its scope to just the outcome wire format and HTTP binding, clearly positioning itself as a profile extension point rather than a complete mechanism. As currently written, it falls uncomfortably between those two options.