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

Review · November 10, 2026

OAuth Client IP Range Metadata, Letting Clients Publish Their Egress IPs to the Authorization Server

draft-mcguinness-oauth-client-ip-ranges -01View on Datatracker →

The problem it's solving

Hosted OAuth clients usually call the token endpoint from a known set of IP addresses. Enterprises want to restrict access to those addresses. Today every vendor publishes its ranges in its own format, and admins copy them by hand into products like Okta network zones or Databricks IP access lists. Those lists go stale when the client's egress changes.

Client ID Metadata Documents make this worse. The AS learns everything about the client from a URL, and there is no field for network origin. A deployment that restricts source addresses has to exempt CIMD clients or refuse them.

The draft's fix is a standard place for the client to declare its IP ranges in client metadata.

How it fits in the working group

This is a small extension that fits the charter. It registers two client metadata parameters and defines one JSON document format. It adds no endpoints, grants, token formats, or AS metadata.

It builds on Dynamic Client Registration (RFC 7591 and RFC 7592) and CIMD. CIMD is listed as an informative reference, but much of the processing is written around it. Prefix syntax is borrowed from the YANG common types in RFC 9911, which is an unusual dependency for an OAuth spec.

It adds one new relationship. The IP range document can be hosted by a third party, such as the client's hosting provider. The AS then trusts that party for the client's network policy.

It overlaps with JAFAR, a format for publishing the IP ranges of automated HTTP clients. This draft defines its own format instead.

What the mechanism actually does

The client declares its ranges one of two ways. client_ip_ranges is an array of IPv4 and IPv6 CIDR prefixes. client_ip_ranges_uri points to an HTTPS document containing the same array. They are mutually exclusive, following the jwks and jwks_uri pattern.

The AS fetches the document and refreshes it on a schedule. It uses normal HTTP caching with ETags. The fetch has SSRF rules: no redirects, no private addresses, bounded size and time.

Declaring a range does not grant access. The AS validates the set, then decides under local policy whether to accept it. Changes can be auto-accepted within pre-approved supernets or held for admin approval.

Failures never widen access. If the document is unreachable or invalid, the AS keeps the last accepted set for a retention period. After that, an AS that was enforcing rejects every request from the client.

Enforcement is optional and AS-only. The AS can hard-block requests from outside the set, or treat them as a risk or rate-limit signal. Resource server enforcement is out of scope.

A match is never authentication. The draft positions this as a supplement to DPoP and mTLS, not a replacement.

So does this stop stolen tokens?

Not as written. The introduction argues that IP ranges catch exfiltrated tokens and DPoP proofs replayed from an attacker's machine. Stolen access tokens are used at the resource server, and the draft defines nothing there.

What it does cover is the token endpoint. Stolen client credentials and refresh tokens stop working from outside the client's network. It does nothing when the attacker shares the same cloud egress as the client, and the draft says so.

Significant interoperability concerns

The client can't tell whether any of this is happening. There is no AS metadata to signal support or a requirement. A DCR client sees the parameter echoed back. A CIMD client gets nothing. Even a registered parameter isn't an accepted one.

There is no error response. This is listed as an open issue. A client rejected for its source address sees a generic failure and can't tell it from bad credentials. Reusing invalid_client could trigger credential rotation for a network problem.

Timing of changes is invisible. The client is told to publish a new prefix before using it. It has no way to know when the AS has fetched it, or whether an admin has approved it. A client that moves egress too early is locked out.

The document host becomes a point of failure. If the range document is down past the retention limit, the client is rejected at every enforcing AS. The retention limit is local to each AS, so the client can't plan for it.

Size limits are unspecified. The AS must bound the number of prefixes, and the draft sets no minimum. A client with 500 prefixes can't know which servers will reject the whole set. With CIMD the rejection is silent.

Empty array handling breaks an RFC 7592 convention. An empty client_ip_ranges means "admit nothing." RFC 7592 treats empty values in an update as a request to delete the field. Existing DCR implementations will get this wrong.

A client can't combine a provider's document with its own prefixes. This is the second open issue. It limits the third-party hosting case the draft calls out.

Two formats for the same data. A company that publishes JAFAR for its crawlers would publish a second document for its OAuth clients. The draft explains how the formats differ, not why the existing one can't be profiled.

Mechanisms that could be removed

  • The retention state machine. Three retention cases, when each timer starts and stops, and narrowing to the intersection of old and new sets are all AS-internal behavior. One rule covers the interop need: a failure or withdrawal must not admit new addresses.
  • The HTTP caching walkthrough. How to handle 304 responses and which ETag to send restates RFC 9111.
  • Source address interpretation. Guidance on X-Forwarded-For and trusted proxies applies to any IP-based control and isn't specific to this metadata.
  • Software statement precedence. RFC 7591 already says software statement values win.
  • Enforcement at the client configuration endpoint. The draft itself notes this can lock a client out of fixing its own declaration.

The interoperable core is two parameters, a prefix syntax, and a one-member JSON document. That would fit in a few pages.

Writing quality

The prose is precise and reads like a mature IETF document. Requirements are scoped carefully and cross-referenced to exact sections of other RFCs. Open issues are listed honestly with links to the tracker.

It is dense for what it defines. This is 23 pages for two metadata parameters. Most of the length goes to AS processing rules that the draft also calls local policy.

Section 4.5.3 on retention is hard to follow. It takes several readings to work out which timer applies in which case.

The motivation and the scope don't line up. The introduction leans on token theft at protected resources, then defers resource servers to a possible companion spec.

Background of the authors

Datatracker could not be reached for this review, so 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. This repo now tracks around 19 distinct McGuinness individual drafts in OAuth.

Xiaoyi Zhang (Databricks). Profile not checked. No other drafts by this author appear in the OAuth, or related working group documents tracked here, which suggests a new IETF author. Databricks is a likely deployer, since its customers already manage IP access lists by hand.

Questions worth raising

  1. Is the AS the right enforcement point? The token theft argument only holds at the resource server. Without that piece, the benefit is limited to the token endpoint.
  2. How does a client learn it was rejected for its IP? The error response needs an answer before adoption, since it decides whether this is debuggable in production.
  3. Should there be AS metadata? A CIMD client has no other way to learn that an AS supports or requires declared ranges.
  4. Why not profile JAFAR? One published range document per operator is better than two.
  5. Can the AS processing rules shrink? The format is the part that needs a standard. Retention and acceptance policy could be a short set of safety rules.

For the competing range format, read draft-illyes-webbotauth-jafar.