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

Review · November 10, 2026

Mission Resource Access, a Generic RAR Type for Resources, Actions, and Constraints

draft-mcguinness-oauth-mission-resource-access -00View on Datatracker →

The problem it's solving

Mission-Bound Authorization records what a user approved for an agent's task as RAR authorization_details entries. It defines no entry type of its own. Every example in it uses mission_resource_access.

This draft defines that type. An entry names one resource, the actions allowed there, machine-checkable constraints such as a spending ceiling, and a policy for which sub-agents can receive it.

Most of the text covers comparison. It says when one entry is a subset of another, so the AS can narrow and delegate. It says how a resource server matches an entry against a request and what it does with parts it doesn't understand.

How it fits in the working group

It is a RAR type definition, which RFC 9396 leaves to whoever defines the API. A generic cross-API type defined by the WG is new territory. RAR deliberately stopped short of one.

The dependency list is short: RAR, Token Exchange, and Protected Resource Metadata, plus URI, JCS, and date-time RFCs. The dependency that matters is the unadopted Mission base draft. It is cited normatively at -01, and the base is already at -02.

The new surface is one RAR type with six members, nine registered constraints, one protected resource metadata parameter, and one new IANA registry. There are two conformance roles and three optional AS capabilities.

It creates no new relationship between roles. It does put new work on the resource server, which has to enforce every constraint in an entry or treat the entry as granting nothing.

Six more Mission companions are cited informatively: runtime, metering, capability binding, AuthZEN, transaction authorization, and cross-domain. They are cited by GitHub URL, not by Datatracker revision.

What the mechanism actually does

An entry is resource, actions, constraints. resource is an absolute URI. resource_match is exact or prefix. actions are strings the resource defines, with wildcard families such as invoices.*. constraints is an object of bounds.

The type replaces RAR's locations with its own resource. An entry carrying locations is refused.

Nine Common Constraints get shared semantics. They are max_amount, resource_issued_after, resource_issued_before, tenant, recipient_domain, time_window, data_classification, allowed_tools, and requires_action_approval. Each has a value syntax, a subset rule, an intersection rule, and a predicate the resource server evaluates. Any other key is deployment-defined.

The AS narrows by subset. A derived entry must have a contained resource, contained actions, every constraint of the parent at an equal or tighter value, and the same contract reference. If the AS can't compare a constraint, it refuses the derivation.

The resource server fails closed. An entry with an unknown member, an unknown resource_match, or a constraint it can't enforce grants nothing. One entry has to authorize the whole request. The resource server can't combine parts of several.

Delegation policy rides in the entry. delegation.max_depth caps the depth. allowed_delegates lists client IDs or actor types such as ai_agent. With no delegation member the entry never appears in a sub-agent's token. The AS enforces this at Token Exchange, and the resource server ignores it.

A contract reference pins meaning. resource_contract carries an ID, version, and SHA-256 digest of a document describing what the resource's actions mean. The AS inserts it from configuration. The resource server refuses the entry unless it performs the operation under exactly that version.

Scope projection lets the AS emit plain scope for resource servers that don't read RAR. Four conditions must hold per entry and per target.

MCP tools map onto it directly. The MCP server URL is the resource. Tool names are the actions. Argument limits are constraints.

So is this a Mission type or a general one?

General, in almost everything but the name. Resource matching, action families, the constraint vocabulary, and the fail-closed rules don't need a Mission. A plain RAR deployment could use them today.

The Mission dependency shows up in delegation, the digest format, scope projection, and the capability declaration. That is the minority of the document.

The WG could look at this without taking a position on Missions. That would need a rename and a normative reference list that stands alone.

Significant interoperability concerns

The semantics live in documents this draft doesn't define. An "operation contract" tells the resource server how to turn a request into a target URI, an action, and constraint inputs. It is described as "documented" and has no format. A resource contract is "a JSON document" with no schema and no publication endpoint. The AS learns the current version from configuration.

Action names are per resource. The AS SHOULD use a namespace the resource documents. An Okta AS can't produce a valid entry for a Salesforce API without Salesforce's action list loaded by hand. This is the same bilateral setup scopes need today.

Two matching modes use two comparison rules. exact is character for character with no normalization. prefix normalizes case, default port, and percent-encoding. The draft's own table shows https://API.EXAMPLE:443/orders/123 failing as exact and matching as prefix. An exact entry for https://api.example doesn't match https://api.example/. The AS has to emit the resource server's exact spelling, which is one more thing learned out of band.

Fail-closed needs discovery, and discovery is optional. If the AS emits a constraint the resource server doesn't enforce, the request gets a 403. mission_constraints_supported in resource metadata is how the AS would know. It is OPTIONAL, and when absent "that knowledge is established out of band."

Several Common Constraints are common in name only. tenant is an unformatted string. data_classification labels are deployment-defined. allowed_tools identifiers have no namespace. Two deployments can compare these identically and still mean different things.

requires_action_approval can't be satisfied from this document. The approval it demands is defined in three companion drafts. A resource server implementing only this draft refuses every such request.

recipient_domain asks a lot of a mail API. The input is "every final recipient of the action, after routing and list expansion." Most senders can't see past a distribution list.

Delegate matching depends on unadopted work. sub_profile comes from the Actor Profile through the base draft. sub matchers are client IDs in one AS's namespace.

The AS keeps hidden state. It must record which approved entries inherited delegate eligibility from policy, and evaluate every later exchange against that record. Nothing on the wire shows it.

Mechanisms that could be removed

  • Resource contracts. A digest over a document with no defined format, published by no defined means. Runtime Enforcement is where the content gets defined, and the member belongs there.
  • Scope projection. Guidance for AS configuration. No protocol element is involved.
  • Transformation Capabilities. Three booleans declared in "deployment documentation."
  • requires_action_approval. Move it to the companion that defines the approval.
  • allowed_tools. Section 12 already models tools as actions. Two ways to name a tool is one too many.
  • data_classification and tenant. Without a shared value space they work as well as deployment-defined keys.
  • Inherited-eligibility provenance and the cross-domain paragraphs. Bookkeeping for the base draft and the cross-domain companion.
  • The locations ban. Either use RAR's common field or explain why resource differs.

What's left is the entry, the two matchers, action families, the subset rule, four or five well-defined constraints, the fail-closed rule, and the metadata parameter. That is a 15-page document, and a useful one.

Writing quality

The draft is precise and in proper IETF style. The matching tables are the best part. They show the percent-encoding and dot-segment cases that path-prefix authorization usually gets wrong, and the security considerations explain each one.

It is long for a type definition at 45 pages. The length comes from serving the base draft's machinery: narrowing, delegation, projection, and capability declaration each get a section.

Cross-references to the base draft are by section name with no numbers, and they are frequent. A reader needs the 166-page base open to follow the delegation and projection sections.

The pinned reference to base -01 is already stale. The document history is a single 150-word sentence.

Background of the author

Sole author, Karl McGuinness, listed as Independent, formerly at Okta.

The Datatracker profile could not be fetched for this review. The figures below are from the profile as checked for the October reviews.

  • 0 published RFCs
  • 16 active Internet-Drafts and 2 expired, as of mid-October 2026
  • Co-author of ID-JAG, an adopted WG document

Not a new author, and writing at a very high rate. The count is higher now. This draft was split out of the Mission base in October, and it cites six further companions.

Questions worth raising

  1. Does the WG want a generic RAR type at all? RFC 9396 left types to API owners. A common resource-and-actions type would help agent deployments, and it is a decision worth making on purpose.
  2. Should it be tied to Missions? Most of the type works without the base draft. Adoption of one shouldn't force the other.
  3. Where do operation contracts and action namespaces get published? Until there is a format and a discovery mechanism, every AS-to-resource pairing is bilateral.
  4. Why not locations? The draft forbids RAR's common field and gives no reason.
  5. Which constraints have real cross-vendor meaning? max_amount, time_window, and recipient_domain are well defined. The rest need a value space or should go.
  6. How does this relate to AuthZEN? Resource, action, and context is the same model. The AuthZEN drafts for claims and issuance are in front of the WG now.

For the framework this type plugs into, read the Mission-Bound Authorization base draft. For a different take on narrowing tool arguments across delegation, read Attenuating Agent Tokens, which this draft cites as possible future work.