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

Review · October 26, 2026

OAuth Scope Metadata Document Summary

draft-forten-oauth-scope-metadata-document -00View on Datatracker →

The problem it's solving

OAuth scopes are opaque strings. When a client or an AS sees a scope it doesn't recognize, nothing tells it what that scope means.

The draft points to two places where this hurts. One is a single AS sitting in front of many resource servers behind a gateway. There, scope names can collide and the AS can't tell which resource defined which scope. The other is an AI agent meeting an API for the first time. It has nothing to go on when it picks which scope to request.

The fix is to make the scope value an HTTPS URL. You fetch the URL and get back a small JSON document describing the scope. The scope is its own discovery mechanism.

How it fits in the working group

It's in charter, and it's small. It defines no new endpoints, parameters, or grant types. It's a convention on top of the scope value from RFC 6749, plus a fetch-and-parse rule.

The design reuses two pieces of existing WG work:

It also sets up a new relationship. Today the AS and client learn about scopes from configuration or developer docs. Under this draft they fetch descriptions at runtime from whoever hosts the scope URL, which is usually the resource server or its gateway.

The normative dependencies are RFC 6749 and CIMD, which is still a draft. Everything else is HTTP and JSON plumbing.

Appendix A gives a fair account of prior art. UMA 1.0 required scope URIs to resolve to a description. UMA 2.0 made that optional because ASes weren't fetching them at runtime. AAuth puts scope descriptions inline in resource metadata.

What the mechanism actually does

  1. The RS uses a scope like https://gw.example.com/scopes/files/write. It has to be HTTPS with a path, no fragment, no userinfo, no dot segments, and ideally no query.
  2. An AS or client that wants to know what the scope means sends a GET to that URL. It may only do this if the URL starts with a "trusted prefix" it has configured.
  3. The response is JSON with an optional description and an optional documentation_uri. Unknown members are ignored.
  4. The AS can show the description on its consent screen. The client can use it to pick scopes.
  5. If the fetch fails, nothing breaks. The scope is treated as an opaque string, as it is today. This is the one deliberate departure from CIMD, where a failed fetch aborts the request.

The document is descriptive only. The AS and RS MUST NOT use it to grant or enforce access.

Significant interoperability concerns

The trusted prefix rule cancels out the main use case. The introduction is about an agent or AS meeting a resource it "has not seen before." Section 4 then forbids fetching any scope URL that isn't under a pre-configured trusted prefix. A resource you've never seen has no trusted prefix. So the headline scenario doesn't work. The author admits this in open issue #3. Until that's solved, this only works for resources someone has already configured. In that case the descriptions could have been configured at the same time.

Existing URL-shaped scopes will get fetched. Google already uses scopes like https://www.googleapis.com/auth/drive. Plenty of enterprise APIs do the same. Under this draft, any URL scope that meets the syntax rules counts as a Scope Metadata URI. A client with a broad trusted prefix will start sending GETs to URLs that were never meant to be fetched and get HTML or 404s back. The trusted prefix limits the damage, but there's no opt-in signal from the RS that says "these scopes are dereferenceable." Something in Protected Resource Metadata (RFC 9728) could provide it.

Consent screens need localization. The main consumer of description is an AS consent screen. A single untagged string won't work for Okta, Microsoft, or Google, which all show consent in the user's language. The draft should support Accept-Language, or language-tagged members like the client_name#ja pattern from Dynamic Client Registration. Otherwise every AS will pick its own approach.

"Descriptive only" sits awkwardly with consent. The draft says the AS must not base authorization on the document. But if the AS shows that text on the consent screen, the user's approval is based on it. A misleading description is a consent phishing vector. The security section says so and then leans on the trusted prefix. The draft should say clearly who is responsible for the text the user sees: the AS operator or the resource host.

There's no guidance on how the client finds out which scopes exist. Dereferencing tells you what a scope means once you have it. An agent picking scopes first needs the list. That comes from scopes_supported in RFC 9728 or the scope parameter in a WWW-Authenticate challenge. The draft should name those paths. It should also explain why per-scope URLs are better than adding descriptions to RFC 9728 metadata, which would take one fetch per resource instead of one per scope.

Mechanisms that could be removed

The prompt injection MUST. "A client based on a large language model MUST treat the description as data, not as instructions" can't be tested and won't change anyone's implementation. It belongs as a SHOULD-level note in Security Considerations, not a normative requirement.

The extra DNS rebinding rule. Checking the connected address, not just the DNS lookup result, is good advice. It applies just as much to CIMD, though. Putting a stricter rule in this draft means two fetch-from-a-URL specs from the same WG behave differently. Better to fix it once in CIMD and reference that.

Possibly the whole document format. With two optional string fields and no localization, the "document" doesn't carry much. If the WG folds scope descriptions into RFC 9728 metadata, which is the approach AAuth already takes, this draft shrinks to the "scope as dereferenceable URL" convention or goes away. That's worth discussing before it grows more members.

Writing quality

Clean and tight. The body runs about three pages, and the IETF language is correct. Each rule borrowed from CIMD is named, and the one place it departs ("unlike [CIMD], which aborts...") is called out explicitly. That makes the draft easy to review.

The open issues are honest and name the right gaps. Issue #3 is the most important one in the document and should be moved into the main text as a known limitation.

It's arguably too short in one place. The introduction describes two use cases but never walks through either one from start to finish. A worked example of an agent hitting a gateway, getting a WWW-Authenticate challenge, fetching scope descriptions, and requesting a token would show quickly whether the trusted prefix model holds up.

Author background

I couldn't reach Datatracker from this session, so this is based on the local draft corpus.

Su-hyeon Forten Lee is the sole author, writing with a personal email address and no listed affiliation. No RFCs by them appear in the corpus, so they look new to the IETF. They're very active right now, with three OAuth individual drafts in October 2026:

Of the three, this is the smallest and most narrowly scoped. It builds on CIMD and RAR metadata rather than adding new token semantics.

Questions worth raising

  1. How does trust work for a resource you've never seen? If the answer is "trust the origin of the protected resource metadata," that should be in -01. Without it, the agent use case doesn't work.
  2. Why per-scope URLs instead of descriptions in RFC 9728? One metadata fetch per resource is cheaper and already fits how MCP clients discover resources.
  3. How does an RS opt in? Existing URL scopes at Google and elsewhere shouldn't start getting fetched just because they're shaped like URLs.
  4. What about localization? A consent-screen feature without language support won't get deployed by large ASes.

For the dereferencing pattern this draft is built on, read Client ID Metadata Document. For the metadata members it borrows, read draft-ietf-oauth-rar-metadata-remediation.