Review · October 26, 2026
OAuth 2.0 Authority Metadata
A scope name tells you nothing about what it actually does. payments.manage might cover reading, creating, executing, or deleting payments. You can't tell from the name.
This draft lets a resource server publish what each scope and RAR type actually permits, as structured JSON. It also lets the resource server say "this scope can be swapped for that smaller one" and "this scope requires a sender-constrained token and phishing-resistant MFA."
The target reader is the authorization server, or a policy engine sitting behind it, deciding whether a request is broader than the task needs. That's the agent use case, but the draft never says "agent." It says "Mission" twice, capitalized and never defined.
How it fits in the working group
It's a metadata extension. It adds one authority_metadata member to both Protected Resource Metadata and AS Metadata. It defines no new endpoints, grants, parameters, or error codes. That fits the charter.
Normative dependencies are light: RFC 6749, 8414, 8707, 9396 and 9728. It also points to the OpenID Foundation AuthZEN token issuance and token exchange profiles as the expected consumers. Resource metadata gets resolved and passed to a policy decision point.
The new relationship is this: the resource server tells the AS how to interpret the scopes the AS issues. Today the AS owns scope definitions and the RS just enforces them. Here the RS publishes its own semantics, and the AS fetches them at request time based on the client's resource parameter. The draft knows this is risky. Most of the security considerations exist to stop an attacker-hosted resource from describing scopes in a way that weakens AS policy.
It overlaps with three existing pieces of work:
- RAR Metadata and Error Remediation, adopted and also by Yaron, already publishes per-type
description,documentation_uriand JSON Schema for RAR types from the AS. This draft publishes a second, different description of the same RAR types. - Protected Resource Metadata already has
dpop_bound_access_tokens_required. The newsender_constrainedboolean says almost the same thing, except it doesn't name a mechanism. - Step-Up Authentication (RFC 9470) already carries
acr_valuesandmax_agefrom the RS. The difference is timing. Step-up sends them in a 401 challenge. This draft publishes them ahead of time, per scope.
What the mechanism actually does
A resource server publishes a list of "authority entries," one per scope or RAR type. Each entry has:
authority: an array of{resource_type, actions, effects, risk}. For example:document, actionsdocument:readanddocument:update, effectsdata-disclosureandstate-change, riskmedium. All of these are free-form strings chosen by the publisher.familyandrank: an advisory ordering.documents.readis rank 10 anddocuments.manageis rank 20 in familydocument_access.access_requirement_profile: a pointer to a named profile carrying token requirements (max_age,max_lifetime,sender_constrained) and authentication requirements (acr_values,max_age).relationships: currently justcan_be_attenuated_to, which points to a smaller entry.
The AS publishes a cut-down version of the same thing: entries with authority, family and rank, but no profiles and no relationships.
In the worked example, a client asks for documents.manage. The AS sees that documents.read is an attenuation target. It decides the task is read-only. It then either issues documents.read or refuses and suggests the smaller scope. The spec doesn't say which, and doesn't define how the suggestion would be sent.
Significant interoperability concerns
Nearly every value is publisher-defined. Action names, effect names, risk labels, families, ranks and profile names are all local to one metadata document. The draft says this repeatedly: "A consumer MUST NOT assume that identical local strings from different publishers have identical semantics." So a generic AS fetching metadata from an RS it has no prior relationship with can't do anything automated with "risk": "medium" or "effects": ["state-change"]. Interop only happens when the AS and RS already share a vocabulary out of band. If they already share a vocabulary, they could share the whole scope definition the same way.
The spec defines no consumer behavior. Section 8 lists "possible uses." Section 8.3 lists six possible attenuation behaviors and requires none of them. Section 6.3 says consumers don't have to follow relationships. Two compliant ASes can read the same document and do completely different things. That can be fine for a descriptive format. But then the spec should say plainly that it is a description format, and stop presenting attenuation as an outcome.
Attenuation has no feedback path to the client. If the AS quietly issues documents.read in place of documents.manage, the client sees a narrower scope in the token response and has to notice it. If the AS refuses, the draft says "No new error code or remediation response is defined here." The client gets a generic error and has no standard way to learn which smaller scope would have worked. The adopted RAR remediation draft has an insufficient_authorization pattern that could carry this. It isn't used.
RAR attenuation is hedged until it means nothing. Sections 6.1, 6.2 and 9.7 each say that for RAR types, can_be_attenuated_to "identifies a candidate" and that the consumer has to work out containment from type semantics on its own. If the consumer needs the type semantics anyway, the relationship adds nothing. Either define the relationship for scopes only, or drop it for RAR.
Two different max_age fields. token.max_age is seconds since the token was issued. authentication.max_age is seconds since the user logged in. Same name, different meanings, sitting next to each other in one profile object. Someone will get this wrong.
AS and RS descriptions can conflict, and the spec only says what not to do. Section 8.2 says the AS must not resolve the conflict "merely by selecting the interpretation that permits issuance." That's a good rule. But it doesn't say which description wins.
Mechanisms that could be removed
- AS-side
authority_metadata. The draft says the AS "can maintain the same information privately without publishing this metadata," and that it doesn't predict issuance decisions. The only consumers named are clients and admin tools. For RAR types this duplicates the adopted RAR metadata draft. For scopes, a simplescope_descriptionsmap would cover the consent-screen case with a lot less machinery. familyandrank. Section 7 lists five things rank MUST NOT be used for. What's left is display ordering.can_be_attenuated_toalready expresses ordering where it matters.effectsandrisk. These are free-form labels with no defined processing. They might drive a policy engine's decisions, but only one that already knows this publisher's vocabulary.sender_constrained. Reusedpop_bound_access_tokens_requiredfrom RFC 9728, or explain why a mechanism-agnostic flag is needed per scope.
What's left is the core idea: per-scope access requirements, and a scope-to-scope "this is a subset" relationship. That's small and useful. It would also fit well as an extension of PRM or of the RAR metadata draft.
Writing quality
The writing is clean and uses proper IETF language. Sentences are short and well formed. The terminology section is good.
The problem is how much of it is disclaimer. A large share of the normative text says what the metadata does not mean: "does not establish," "does not imply," "MUST NOT be interpreted as," "does not prove." Section 8.1 is entirely negative. Section 7 lists five prohibitions and three permissions. The same RAR caveat appears in Sections 3.1.6, 4, 4.2, 6.1 (twice), 6.2 and 9.7. Cutting the repeats would shrink the document by about a fifth without losing anything.
So much hedging is itself a signal. The author seems well aware that the structured data can't safely drive automated decisions without shared vocabularies. The draft would be stronger if it said that once, up front, and then said what the data is good for.
Smaller issues:
- "Mission" is capitalized as if it refers to a defined concept from another spec, but nothing is cited.
- RFC 9068, 8705 and 9449 are listed as informative references but never cited in the text.
- Acknowledgments is still "TODO."
- Appendix A already flags that extension registries for kinds, relations and requirement attributes may be needed. They are. Without them, "unrecognized kind" has nowhere to come from.
About the author
Yaron Zehavi, Raiffeisen Bank International. Datatracker couldn't be reached for this post. Per our review of his audience-bound token exchange draft two weeks ago, his profile showed zero published RFCs and four active drafts. One of those, RAR Metadata and Error Remediation, has been adopted by the working group. Several other drafts have expired.
So not a new author. He's one of the most active individual contributors in the OAuth WG right now. This is his fifth active OAuth draft, and it overlaps with the one the WG already adopted.
So what should happen next?
The problem is real. Opaque scope names are a known pain point, especially as agents start requesting access on someone's behalf and the AS needs to judge whether the request is broader than the task.
The most useful question for the mailing list is whether this should be a standalone spec at all. The per-scope access requirements and the scope-to-scope attenuation relationship are the parts that would actually change behavior. Both would fit as additions to RAR Metadata and Error Remediation, which already has a remediation path that could tell the client what smaller request would succeed. The free-form actions, effects and risk vocabulary is better worked out alongside AuthZEN, where policy engines actually consume it.