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

Review · November 10, 2026

RFC 9728bis, Relaxing the Protected Resource Metadata Resource Match to a Path Prefix

draft-mcguinness-oauth-rfc9728bis -02View on Datatracker →

The problem it's solving

Protected Resource Metadata has a strict rule. When a client follows a WWW-Authenticate challenge to a metadata document, the resource value in it must be identical to the URL the client just requested.

That rule breaks down for real APIs. A request to https://example.com/users/100 needs metadata whose resource is exactly https://example.com/users/100. The server either publishes a metadata document per URL, or the client has to know out of band which "parent" URL to hit first.

The client also can't tell that a token for /transactions is good for /accounts. It either repeats discovery for every URL or reuses the token on a guess.

The draft's fix is one rule change. The resource value can be any URL on the same origin whose path is a prefix of the request URL.

How it fits in the working group

This is a targeted update to a single paragraph of RFC 9728 Section 3.3. It fits the charter and carries an "Updates: 9728" header.

It adds no endpoints, parameters, metadata fields, roles, or IANA registrations. The well-known retrieval rule is untouched. Metadata fetched directly from a well-known URL still requires an exact match.

It depends on RFC 9728, RFC 8707, RFC 3986, and RFC 9110. The change is backwards compatible in one direction: metadata that passes the exact-match rule also passes the new rule.

-02 has no changes from -01. It is a refresh to keep the draft from expiring.

So how does this relate to MCP? MCP clients discover everything through the WWW-Authenticate challenge, and the acknowledgments credit the MCP Authorization Working Group. This is the population that hits the exact-match rule most often.

What the mechanism actually does

The client runs four checks instead of one. The resource value must be https. Its host must match the request URL's host. Its port must match, with 443 as the default. Its path must be a prefix of the request URL's path.

Prefix matching happens on segment boundaries. /api/v1 covers /api/v1/accounts. It does not cover /api/v10 or /api/v1admin. A path of / covers the whole origin.

The resource server picks how wide the resource is. It can advertise the whole host, a path like /api/v1, or keep using the exact URL.

The client uses the advertised resource value in its token request. That value goes to the AS as the Resource Indicators parameter. The AS only needs to know one identifier per API instead of every URL.

The client can cache tokens by AS and resource. A token obtained for https://example.com/users/ can be reused for any URL under that prefix without repeating discovery. This part is non-normative guidance.

Significant interoperability concerns

Any path can claim its ancestors, including the whole origin. A request to https://shared.example.com/tenantA/x can return metadata with resource set to https://shared.example.com/. That passes all four checks. The client then asks the AS for a token for the entire origin and sends it to tenant A. Section 7.2 says path-based multi-tenant hosting is supported. Section 7.5 says everything on an origin shares one trust boundary. Both can't be true. The exact-match rule limited a resource to speaking for its own URL, and this is the main security property the update gives up.

Token caching turns that into token leakage. A client that cached a token for / will send it to every path on the origin, including paths that never issued a challenge. The draft says the client "should not assume" all resources under a prefix accept the same token, then tells it to try anyway.

Decoding percent-encoded characters before comparison is wrong. RFC 3986 normalization only decodes unreserved characters. Decoding %2F turns one segment into two, which changes where segment boundaries fall. Two clients following different readings will disagree on whether /api%2Fv1/x matches /api/v1.

Query strings are not addressed. RFC 9728 allows a resource identifier with a query component. The request URL almost always has one. The draft only defines path comparison and says nothing about what to do with either query.

Which value goes in the token request is never stated normatively. Under the old rule the request URL and the resource value were the same string. Now they differ. Only the non-normative appendix says to send the metadata value.

The examples don't line up with the well-known rule. Appendix A serves resource: https://example.com/users/ from /.well-known/oauth-protected-resource. A client deriving the well-known URL from that resource identifier would look under /.well-known/oauth-protected-resource/users/. The document is valid when reached by challenge and invalid when reached directly.

Old clients reject new metadata. A deployed RFC 9728 client enforces exact match. A resource server that switches to a prefix-style resource breaks every one of them, and has no way to tell which rule a client follows.

Mechanisms that could be removed

  • The token caching section. It is labeled non-normative but uses MAY. It is also where the leakage risk comes from. Either make it normative with real constraints or cut it.
  • The sender-constrained tokens section. It says DPoP is unaffected. That is one sentence, and it pulls in RFC 9449 as a normative reference.
  • One of the two motivating examples. The introduction and Appendix A make the same argument with different URLs.
  • The Examples section or the match table. The table already shows the algorithm. Section 5 repeats it in prose.
  • Unused references. RFC 6749, RFC 6750, and the resource-token-response draft are listed and never cited.

Writing quality

The prose is clean and uses standard IETF style. The "original rule, updated rule" layout makes the change easy to review. The match table is the most useful part of the document.

At 16 pages it is about twice as long as a one-paragraph update needs. The normative content is under two pages.

There are small slips. "instaed" is a typo in Appendix A. The "TLS origin" definition cites Section 4.3.2 of RFC 9110, which covers http origins. The https origin text is in 4.3.3. The reference to the resource-token-response draft points at -03, and -04 is out.

The security considerations need the most work. The upward-claim case is waved off as "no different from the existing web security model" without walking through what the client does with the resulting token.

Background of the authors

Datatracker could not be reached for this review, so current profile counts are not verified.

Karl McGuinness (Independent, formerly Okta). As of the mid-October reviews: 0 published RFCs, 16 active Internet-Drafts, 2 expired. Co-author of ID-JAG, an adopted WG document. Not a new author, and writing at a very high rate.

Aaron Parecki (Okta). Profile not checked. Co-author of RFC 9728, the document this draft updates, and editor of OAuth 2.1. An established author with several active WG drafts.

Questions worth raising

  1. Should a resource be able to claim a path above itself? Without some limit, the multi-tenant story in Section 7.2 doesn't hold. Options include requiring the metadata URL to sit at the well-known location derived from the resource value, or having the client confirm the claimed prefix independently.
  2. What is the comparison algorithm, exactly? Pin it to RFC 3986 Section 6.2.2 and 6.2.3, drop the blanket percent-decoding, and say what happens to query components.
  3. Should the well-known rule and the challenge rule agree? Requiring the challenge to point at the well-known URL for the advertised resource would make both retrieval paths give the same answer.
  4. How does a resource server migrate? Existing RFC 9728 clients reject prefix-style metadata. The draft should say so and give deployment guidance.
  5. Is this the fix, or is the token response the fix? Returning resource from the AS tells the client what a token covers without loosening the metadata check. The two drafts should state how they work together.

For the companion mechanism on the AS side, read draft-mcguinness-oauth-resource-token-resp.