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

Review · August 20, 2026

Agent Authorization Profile (AAP) for OAuth 2.0

draft-aap-oauth-profile -01View on Datatracker →

Agent Authorization Profile (AAP) for OAuth 2.0

Draft: draft-aap-oauth-profile-01 Author: A. Cruz (Independent)


Problem Statement

Standard OAuth 2.0 was designed for user-delegated authorization, where a human approves a scope at login time and an app acts on their behalf. Autonomous AI agents break this model in several ways: they act at high frequency without human presence, risk levels vary depending on what task they are executing, and they may need to delegate authority further to sub-agents or tools. The spec argues that scope strings alone are too coarse-grained to capture these nuances safely and proposes a structured JWT profile to address them.


How the Spec Fits In

This spec sits squarely in the space of OAuth access token profiles, layered on top of:

  • RFC 6749 — OAuth 2.0 core (Client Credentials Grant)
  • RFC 7519 — JWT
  • RFC 8693 — Token Exchange (for delegation)
  • RFC 9449 — DPoP proof-of-possession
  • RFC 8705 — mTLS certificate-bound access tokens
  • RFC 7009 — Token Revocation
  • RFC 7662 — Token Introspection

Rather than defining a new grant type or protocol flow, it specifies a set of structured JWT claims that Authorization Servers MUST populate and Resource Servers MUST validate. The seven new top-level claim objects are: agent, task, capabilities, delegation, oversight, context, and audit.

This makes it a relatively self-contained profile, similar in spirit to how GNAP or transaction tokens extend existing OAuth infrastructure, but targeting AI agent use cases specifically.

A notable potential conflict exists with related ongoing work:

  • draft-ietf-oauth-transaction-tokens — already addresses short-lived context propagation tokens in workload/microservice scenarios. There is conceptual overlap with AAP's task and audit bindings.
  • WIMSE (Workload Identity in Multi-System Environments) — the IETF working group specifically addressing agent and workload identity. AAP's agent identity claims may duplicate or conflict with what WIMSE standardizes.
  • GNAP — includes rich delegation semantics. The delegation chain in AAP echoes GNAP concepts without reference to them.

The spec does not discuss or cite any of these adjacent efforts, which is a significant gap.


High-Level Summary

AAP tokens are JWTs that carry structured authorization context specifically suited for AI agents. Here is what each claim object provides:

agent — Identifies the software agent: its ID, type (e.g., llm-autonomous), the operator organization that registered it, the model and version it runs on, and whether its runtime environment has been attested. This gives Resource Servers a rich identity picture beyond a simple client ID.

task — Binds the token to a specific unit of work: a task ID, declared purpose, data sensitivity level, and who initiated it. This allows an RS to reject requests that seem inconsistent with the declared task.

capabilities — Replaces or supplements OAuth scopes with a list of named actions (using a dot-namespaced grammar, e.g., search.web, cms.publish) each accompanied by structured constraints: rate limits, allowed/blocked domains, CIDR ranges, HTTP method allowlists, data classification maximums, and geographic restrictions. This is the most substantive new mechanism in the spec.

delegation — Tracks how many Token Exchange hops have occurred, a maximum allowed depth, and the full chain of agent IDs that have been involved. An AS issuing a delegated token must never grant capabilities the parent token did not have.

oversight — Lists which capability actions require explicit human approval before execution and references an approval policy. The spec does not define how that approval is obtained or signaled — it is purely a flag the RS checks.

context — Restricts when and where a token is valid: network zones, time windows with clock skew tolerance.

audit — Carries log level, a trace ID, and a session ID for end-to-end traceability across multi-agent workflows.

The spec prescribes a 12-step RS validation sequence, defines custom OAuth error codes (e.g., aap_invalid_capability, aap_approval_required), specifies token lifetime guidelines by risk level (5–120 minutes depending on action sensitivity), and includes a migration path from scope-based to capabilities-based authorization.


Writing Quality

The draft shows familiarity with IETF conventions: it uses RFC 2119 normative language (MUST, SHOULD, MAY), includes ABNF grammar for the action-name format, registers claims with IANA, provides a threat model section, and separates security and privacy considerations. These are all positive signs.

However, there are notable concerns:

Scope creep / excessive prescription. The spec defines specific token lifetime values (60 min for read-only, 5 min for execute), a "refresh at 80% of lifetime" heuristic, a "reject tokens larger than 16KB" requirement, 90-day key rotation schedules, and HSM storage recommendations. These operational guidelines belong in a Best Current Practice (BCP) document, not a protocol spec. Their presence inflates the document and introduces implementation details that may not be appropriate to standardize.

Missing justification. Several design choices are stated but not justified. Why must rate-limiting windows reset at the clock hour rather than use a sliding window? Why is data_classification_max a total ordering across four levels? Why is the delegation chain validated by length equality rather than content? These deserve explanation.

Incomplete mechanism for human oversight. The oversight claim flags which actions require human approval, but the spec explicitly says nothing about how approval is obtained or communicated back to the token. This leaves the most important safety mechanism (human-in-the-loop) entirely undefined.

No protocol flow diagrams or examples. An end-to-end example showing the AS, agent, resource server, and a sub-agent delegation would significantly improve readability.

Privacy section is thoughtful but detached. The guidance on data minimization and cross-domain correlation is well-reasoned, but most of it is advisory. Very little of it rises to normative requirements.

Overall the writing is competent and the structure is sound, but the document is trying to do too much for a first individual draft, and a substantial portion of its content belongs in separate operational or guidance documents.


Author Background

Angel Cruz ([email protected]) — Independent

  • RFCs published: 0
  • Active Internet-Drafts: 0
  • IETF experience: This draft is their sole recorded contribution to the IETF.

Angel Cruz is a new entrant to the IETF process with no prior RFCs and no other active drafts. This is their first submission. That is not a disqualifier — good ideas come from newcomers — but it does mean the draft has not had the benefit of sustained engagement with the working group community, and the absence of co-authors with IETF experience may explain some of the structural and scoping issues noted above.


Suggested Feedback

  1. Engage with WIMSE and transaction-tokens work. The authors should study draft-ietf-oauth-transaction-tokens and the WIMSE working group output and either differentiate from or align with those efforts. Submitting an OAuth profile that duplicates agent identity work already underway without acknowledgment will be a red flag for the WG.

  2. Narrow the scope. Strip out operational recommendations (token lifetimes, key rotation schedules, HSM guidance, token size limits) into a separate informational or BCP document. The normative spec should focus on the claim structure and validation semantics.

  3. Define or reference the human oversight mechanism. The oversight claim is the most important new concept for safety, but the spec leaves it unimplemented. Either define a minimal signaling mechanism or drop the claim and leave it for a follow-on draft.

  4. Add co-authors with IETF experience. Working with seasoned IETF authors would significantly improve the document's chances of adoption and help navigate working group process.

  5. Clarify the relationship to scopes. The spec says both scope and capabilities may be present and that AAP-aware RSes must use capabilities as authoritative. This creates an ambiguity for RSes that are partially aware. The interaction model needs to be specified more precisely.