Review · August 26, 2026
SCIM Extension for Tenant-Aware Identity Provisioning
draft-kushwaha-scim-tenant-resource-00 — Saurabh Kushwaha (Oracle), submitted 25 August 2026. SCIM working group, individual submission, 29 pages.
The Problem Statement
Every multi-tenant SCIM service provider has a tenant concept. None of them are the same concept, because SCIM never defined one. So "tenant," "organization," "identity domain," and "workspace" are all the same idea with four different shapes on the wire, and nothing portable connects them.
The draft's second problem is the more interesting one. A SCIM request can carry tenant intent in several places at once: a claim in the access token, a segment in the URL, an HTTP header, the request's subdomain. When two of them disagree, which one wins? Nothing says. If the answer is "the URL," an attacker with a token for tenant B can point the URL at tenant A and see what happens.
That's a real attack, and it's the part of this draft worth the working group's time.
How the Spec Fits In
This is four schema URIs and a set of protocol rules:
core:2.0:Tenant, a new resource type at/Tenantsextension:tenantMembership:2.0:User, a multi-valued membership arrayextension:tenantMembership:2.0:Group, tenant ownership for groupsextension:multiTenancy:2.0:ServiceProviderConfig, capability discovery
The URNs are correctly formed, which is worth saying because the author's agent governance draft got this backwards. Putting the new resource type in the core namespace matches how RFC 9944 registered core:2.0:Device and core:2.0:EndpointApp. That choice has a consequence the draft doesn't mention: RFC 7643 §10.3.1 says registration of new core schemas MUST be reviewed by the designated expert and published in an RFC. Unlike the agent governance extension, this document genuinely has to be an RFC.
The draft cites the wrong section of RFC 7644, and the right one makes its case better.
The introduction says RFC 7644 §1.1 "acknowledges that a service provider may be multi-tenant and lists deployment techniques." RFC 7644 §1.1 is "Intended Audience." The content the draft is describing is Section 6, "Multi-Tenancy," three pages with two subsections. And §6 says considerably more than the draft credits it with:
- §6.1 lists the three explicit tenant specification methods: URL prefix, sub-domain, HTTP header. It even names the placeholder
{tenant_id}. - §6.1 says, in 2015: "The service provider should consider precedence in cases where a client may explicitly specify a Tenant while being implicitly associated with a different Tenant." That is exactly the gap §8 of this draft fills. The draft is answering a question RFC 7644 asked and left open, which is a much stronger pitch than "SCIM doesn't mention this."
- §6.2 says externalId is unique only within the tenant. §12 of this draft calls that "long-standing precedent in SCIM." It's not precedent, it's normative text in the protocol RFC.
Fixing the citation is a one-line change that materially improves the draft's argument.
Two of the draft's premises are already answered in the core RFCs.
RFC 7643 §2.2 defines uniqueness: server as "The value SHOULD be unique within the context of the current SCIM endpoint (or tenancy)." Tenant-scoped userName uniqueness is already conformant with SCIM 2.0 today. §12 frames itself as specifying "how tenant-scoped uniqueness interacts with the uniqueness = server constraint," but there's no conflict to resolve. The real gap is that a client can't discover which scope a server chose, and the tenantScopedUniqueness config object solves that. Narrow §12 to the discovery problem and it gets shorter and more defensible.
Going the other direction, RFC 7644 §6.2 says "The service provider may choose to implement SCIM ids that are unique across all resources for all Tenants, but this is not required." §12 item 3 of this draft says id and meta.location "remain server-unique" and that the document "SHALL NOT weaken id uniqueness." That's a new constraint on deployments RFC 7644 explicitly permits. A provider using tenant-prefixed base URLs with per-tenant id counters is conformant today and non-conformant under this draft. The draft claims backward compatibility in its abstract; this is the one place it isn't.
Charter fit is partial, and it splits along the seams of the document.
Multi-tenancy is not a listed item in the SCIM charter. Two pieces of this draft do have clean hooks:
- Tenant-scoped uniqueness discovery lands on "Updates to the evolution of the externalid usage."
- The OAuth token claim lands on "Profiling SCIM relationships with other identity-centric protocols such as OAuth 2.0."
The Tenant resource type, the lifecycle state machine, and the region profile have no hook other than the general clause about "extensions to SCIM that have found broad adoption." The draft's own premise is that no portable version of this has been adopted by anyone, so that clause doesn't carry it. This is answerable, but the chairs should answer it before the document grows further.
The roles delegation rests on something that isn't there yet. §1.1 and §5.1 say tenant-local roles are deliberately excluded because RoleAssignment handles them, and that RoleAssignment's scope type "tenant" "references a Tenant defined by this document." Checking draft-poreddy-scim-role-assignment: "tenant" appears in a list of common example values with no canonicalValues array, and scope.value is described as a "Provider-meaningful identifier." It does not reference a SCIM Tenant id. §18 of this draft asks whether Tenant "should be a registered RoleAssignment scope type," which is the honest version. §1.1 and §5.1 state it as settled. Cutting scope by pointing at an unadopted single-author draft is the right instinct and the wrong load to put on that draft.
Ten normative references, three of them drafts on the informative side.
High-Level Summary
Picture an identity governance platform that provisions into a SaaS product for 400 customer tenants. One OAuth client, one set of credentials, 400 tenants. Every request has to say which tenant it means.
Today the SaaS product picks a mechanism, documents it in a developer portal, and the governance platform writes tenant-selection code specific to that vendor. Do that 30 times and you have 30 code paths. That's the portability problem.
The security problem is sharper. Say the product accepts the tenant in the URL path and also in the access token. An attacker holds a legitimate token for tenant-b and sends PATCH /Tenants/tenant-a/v2/Users/.... If the server reads the path and authorizes off the token, or reads one and forgets to check the other, the write lands in the wrong customer's tenant. Nothing in SCIM 2.0 tells the implementer which to trust.
§8 is the answer and it's the best-constructed part of the draft. It defines the effective tenant T* as the highest-precedence authenticated indicator, where "authenticated" defaults to the access token claim only. Then it requires two things before anything commits: the client must be authorized for T*, and every other indicator in the request must resolve to the same T*. The second condition is what stops a spoofable indicator from quietly redirecting a request that passed authorization on a different basis.
The draft then does something most -00 drafts don't. It orders the two checks so that the conflict error (400) is only reachable after the authorization gate has passed. That means an attacker can't use the difference between "conflict" and "not found" to probe which tenants exist. Unauthorized and nonexistent both return an identical 404. The reasoning is written out in §8.2 and §15.2, and it's correct.
For a leader evaluating this, the summary is: one good idea, three documents.
The tenant resolution rule is a genuine contribution that closes a question the working group left open in 2015. Around it sit a full resource type with lifecycle state machine, a uniqueness and compatibility contract, and an optional multi-region data-residency profile with GDPR transfer guidance. Those are three separate pieces of work with three different sets of reviewers and three different charter arguments. Bundled together they're hard to review and easy to stall.
The operational catch: §10.1 requires ETag support. RFC 7644 §3.14 makes ETags a MAY, and a large share of deployed SCIM servers don't implement them. Turning an optional core protocol feature into a conformance requirement is the highest-cost line in this document, and it's one sentence in a subsection.
The second catch: the configurability. A conformant server can accept any subset of five indicator methods, in any precedence order it declares, with any subset treated as authenticated, under a configurable claim name and a configurable header name. Every one of those is advertised, so a client can in principle read the config and adapt. In practice no client will implement five methods and a runtime-ordered precedence walk. They'll implement whatever their two biggest partners do.
Significant Interoperability Concerns
Two of the five indicator methods have no wire format. §8.3 says a path or body indicator "carries a Tenant id directly" and stops there. Where in the path? RFC 7644 §6.1 shows https://example.com/Tenants/{tenant_id}/v2/Users, but this draft never says. Which body field? Never named. Both appear in tenantSelectionMethods and both are in the default tenantSelectionPrecedence. You cannot implement either one interoperably from this document, yet a server can advertise support for both and be conformant.
scimType is paired with 409 in four places, and RFC 7644 doesn't allow that. RFC 7644 §3.12 introduces the detail error types with "For HTTP status code 400 (Bad Request) responses, the following detail error types are defined," then lists invalidFilter, tooMany, uniqueness, mutability, invalidSyntax, and the rest. The draft uses HTTP 409 with scimType = mutability in §9.1 and §13.2, and HTTP 409 with scimType = uniqueness in §9.3 and §10.2. Those pairings aren't defined. Either use 400, or say plainly that the document extends the applicability of those codes to 409 and register the extension.
uniqueness is the wrong code for the cascade refusal regardless of status. §9.3 rejects a tenant deletion that would orphan identities with scimType = uniqueness, which RFC 7644 defines as "One or more of the attribute values are already in use or are reserved." Refusing a delete because dependents exist is not a uniqueness violation. This is the kind of thing that gets copied into an implementation and then into a test suite.
404 for unauthorized overrides RFC 7644 Table 8 without saying so. Table 8 assigns 403 to "Operation is not permitted based on the supplied authorization." §8.2 and §15.4 require 404 instead, deliberately, for enumeration resistance. That's the right security call and it's a documented override of the protocol RFC's status code table. The draft should say it's an override. Right now a reader has to notice.
Creating the first membership is impossible as written. §5.2 item 4: "If no membership exists for t, the user has no presence in t, and SCIM operations targeting t for that user SHALL be rejected with HTTP 404." A PATCH that adds the first membership for a user in tenant t is an operation targeting t for a user with no membership in t. Read literally it gets a 404 forever. §5.2 needs to carve out membership-creating operations.
Creating the first tenant has the same problem. §14: "MUST resolve each request to a single effective Tenant T* per Section 8." A POST /Tenants has no tenant to resolve to. Both bugs come from applying the resolution rule to every request without exempting the operations that bootstrap the thing being resolved. Fixing them together is a scoping statement in §8: the rule applies to operations on tenant-scoped resources, and /Tenants collection operations and membership creation are governed separately.
§9.1 and §12 contradict each other on migration. §12 says a deployment migrating from server to tenant uniqueness "SHOULD perform the migration while the affected Tenants are in status suspended." §9.1 says that while a tenant is suspended, the service provider "SHALL reject state-changing operations on resources owned by that Tenant." A migration is a state-changing operation on resources owned by that tenant. One of these has to give.
The Group extension's attribute is named value. So a group carries "urn:...:tenantMembership:2.0:Group": {"value": "tenant-acme", "scope": "tenant"}, and the §11 filter example reads urn:...:tenantMembership:2.0:Group:value eq "tenant-acme". Group.members.value already exists in RFC 7643 and means a member's id. Anyone reading that filter will read it as a member query. Name it tenant or tenantId.
While you're there: the Group side isn't membership, it's ownership. One tenant owns the group, per §6.2. The User side is a many-to-many array called memberships. Same extension name, two different relationships, two different shapes. Call the Group one what it is.
Two status vocabularies that overlap but don't match. Tenant status is active / suspended / archived / pending. Membership status is active / suspended / invited / pending / deactivated. Same attribute name, two schemas, three shared values, and archived versus deactivated as the respective end states. An implementer will write one enum and reuse it. Consider making them visibly distinct, or explain in §9 why they aren't.
The JWT claim registration asks for a status that doesn't exist. §17.3 requests tenant_id "as provisional" per RFC 7519 §10.1. RFC 7519 §10.1 is Specification Required with a three-week review on [email protected]. There is no provisional state in the JWT Claims registry. HTTP field names do have one, so §17.4 is fine; §17.3 isn't.
The substance behind it is worth more attention than the procedure. tenant_id would be a general-purpose OAuth access token claim minted by a SCIM document. RFC 7519 §10.1 tells the designated experts to weigh "whether it is likely to be of general applicability or whether it is useful only for a single application." A tenant identifier claim is broadly applicable, which is an argument for registering it and also an argument that the OAuth working group should be in the conversation. §18 raises this. Raise it on both lists.
The ServiceProviderConfig shape diverges from the working group's own recent precedent. RFC 9865, published October 2025, added a pagination complex attribute directly to the ServiceProviderConfig document under the base core:2.0:ServiceProviderConfig schema URN, marked OPTIONAL, with most sub-attributes OPTIONAL. This draft defines a fourth schema URI, nests everything one level deeper, and marks the container Required. Both work. Only one matches what the WG did nine months ago.
§13.2 cites RFC 7644 §4 for Link header semantics. RFC 7644 §4 is "Service Provider Configuration Endpoints." It has no Link header semantics. If the region profile survives, this needs RFC 8288.
Mechanisms That Could Be Removed
The draft asks in §18 whether the region profile should be a companion document. Yes. §13, plus §15.8, plus §16.3 with its GDPR Article 44 through 49 citations, is a data-residency document wearing a tenancy document's clothes. It's the only part of the draft that requires reviewers who know cross-border transfer law, and holding the rest hostage to that review is a bad trade. Cite regional statutes sparingly in the split-out version too; RFC practice is to describe the requirement, not the Article number.
Drop the subdomain and body indicators. This is the highest-value cut. Removing subdomain also removes FQDN canonicalization, trailing-dot handling, IDN A-label conversion, the domains attribute with its verified sub-attribute, and a third of §8.3. Removing body removes an indicator that was never given a field name. That leaves tokenClaim, path, and header, which is exactly the set RFC 7644 §6.1 already named, minus subdomain. Three methods with a fixed precedence is implementable. Five with a reorderable precedence is a menu.
Make tenantSelectionPrecedence fixed. The authenticatedSelectionMethods knob already covers the legitimate variation, and §8.7 plus §15.3 spend most of a page constraining the reordering to the point where the only permitted deviations don't change behavior. If the token claim always wins when present, say the token claim always wins when present, and delete the ordered list, the deviation rules, and the paragraph justifying them.
§14, "Normative Behavior Summary," should go or stop being normative. Every MUST in it restates a requirement from §§7 through 13. A conformance summary that duplicates normative text is a place for the two copies to drift, and the draft already has one instance of that in the §9.1 / §12 contradiction above. Keep it as a non-normative table if it's useful for reviewers.
§15.2, "Why the Rules Hold," is argument, not specification. It's good argument and it belongs in the working group discussion. In a published RFC it reads as a proof sketch that isn't a proof, which the section itself concedes. Compress to two sentences or move to an appendix.
Tenant.type has no interoperability content. Five canonical values (customer, partner, internal, workspace, organization) and no behavior anywhere in the document keys off any of them. Either something depends on it or it's a label, and labels don't need canonical values.
§16.2 conflicts with the schema model. It says /Tenants list responses "SHOULD by default return only id, displayName, and status." RFC 7643 already expresses per-attribute return behavior through the returned characteristic. Set returned on the attributes in §4.2 and delete the prose.
Writing Quality
This is a well-constructed document, and the technical writing in §8 is better than most of what arrives at -00. The resolution rule is stated in set notation with the invariant named separately, the error ordering argument is explicit, and canonicalization is handled rather than assumed. That's careful work.
The problems are structural and mechanical rather than a matter of clarity.
BCP 14 usage is inconsistent within the document. §§4 through 13 use SHALL almost exclusively. §14 and §15 switch to MUST. The SCIM RFCs use MUST. Pick one.
"Terminal-soft state" is coined in Table 1 and never defined. The reader has to infer it from the transition table.
The same argument appears three times. The closing paragraph of §8.1, the bullets in §14, and the prose in §15.2 all walk through why the two-step check preserves the invariant. Once is enough.
The Change Log describes changes "relative to the author's working draft." For a -00 there's nothing public to diff against, so the log is nine bullets of revision history for a document nobody has seen. It reads oddly, and it says something about process: this draft was iterated privately rather than on the list. Compare the author's agent governance draft, whose acknowledgments credit Danny Zollner and Pamela Dingle for specific feedback that visibly changed the scope. Here the acknowledgments thank the working group generically and name no reviewers.
On length: 29 pages, and roughly 12 of them are the tenant resolution rule and the schemas it operates on. The rest is a lifecycle state machine, a uniqueness contract, a region profile, a conformance summary that repeats the body, and a security section that argues its own correctness. Split three ways, each piece would be short enough to actually get reviewed.
Background of the Author
Saurabh Kushwaha (Oracle, [email protected]) — 0 RFCs, 4 active Internet-Drafts, 0 expired. No IETF group roles. New IETF author, writing a lot right now.
All four active drafts are SCIM, all individual submissions:
draft-kushwaha-scim-tenant-resource(this one)draft-kushwaha-scim-agent-governancedraft-kushwaha-scim-didvc-binding(at -01)draft-kushwaha-scim-attr-cursor-pagination
This is the fourth of four and by a wide margin the most invasive. The other three are extension schemas. This one adds a resource type, mandates request-processing behavior, overrides RFC 7644's status code table, promotes an optional core protocol feature to mandatory, and requests registrations in two registries outside SCIM's control. Four individual drafts into one working group from a participant with no RFCs is already a lot of review surface. One of them being a protocol profile makes the sequencing question sharper: which of the four does the author most want adopted?
Worth noting in the author's favor: the URN construction is correct here, the core namespace choice matches RFC 9944, and §18 lists six open issues honestly, including the two structural ones a reviewer would raise first. The affiliation is Oracle and the address is an Oracle address, unlike the agent governance draft's personal account.
Suggested Feedback
Fix the RFC 7644 citation, because §6 helps you. §6.1 already lists the three indicator methods and already says the service provider "should consider precedence" when explicit and implicit tenant association disagree. Pointing at that turns the draft's pitch from "SCIM is silent" into "SCIM named this open question in 2015 and here's the answer." It's a stronger argument and a one-line fix.
Split it three ways. Tenant resource plus membership plus resolution rule in one document. Tenant-scoped uniqueness discovery in a second, which has a clean charter hook on externalId evolution. Region profile in a third. §18 already asks about the third; the answer is yes, and the second should go with it.
Two of the five indicators can't be implemented.
pathandbodyhave no wire format anywhere in the document. Either specify them or drop them. Recommend droppingbodyandsubdomainentirely and keeping the three RFC 7644 §6.1 already names.The configurability is the main interop risk. Five methods, reorderable precedence, configurable authenticated subset, configurable claim name, configurable header name. Advertised or not, no client will implement the full matrix. Fix the precedence, keep
authenticatedSelectionMethods, and the spec gets shorter and more likely to interoperate.scimTypewith 409 isn't defined. RFC 7644 §3.12 defines those detail codes for 400 responses only. Four places in this draft pair them with 409, anduniquenessfor a cascade-delete refusal is semantically wrong on top of that.Two bootstrap bugs, one cause. §5.2 item 4 makes the first membership uncreatable and §14's universal resolution requirement makes the first tenant uncreatable. Both need §8 scoped to operations on tenant-scoped resources.
ETags as SHALL is the biggest adoption cost in the document. RFC 7644 §3.14 makes them MAY and plenty of shipping servers skip them. If optimistic concurrency is required for membership mutation specifically, say that and explain why the rest of the resource doesn't need it.
RFC 7643 §2.2 already says "or tenancy."
uniqueness: serveris defined as unique "within the context of the current SCIM endpoint (or tenancy)," so tenant-scopeduserNameis conformant today. §12 should be about discovery, not about resolving a conflict that isn't there. And going the other way, §12's requirement thatidbe server-unique contradicts RFC 7644 §6.2, which explicitly permits per-tenant ids. That's the one place the backward compatibility claim doesn't hold.Take the JWT claim to the OAuth list. There's no provisional status in the JWT Claims registry, so §17.3 needs rewriting regardless. More to the point,
tenant_idis a general-purpose access token claim, and RFC 7519 §10.1 directs the designated experts to weigh general applicability. That's an argument for registering it and an argument for the OAuth working group being in the room.Don't lean on RoleAssignment yet. Removing tenant-local roles was the right call. But RoleAssignment lists "tenant" as an illustrative scope type with no canonical values, and its
scope.valueis "provider-meaningful," not a SCIM Tenant id. §1.1 and §5.1 state the composition as settled; §18 admits it isn't. Match the two, and say plainly that scoped roles are deferred rather than solved elsewhere.
On the merits: §8 is a real contribution and the working group should want it. The tenant resolution rule answers a question RFC 7644 left open, the error-ordering analysis is correct, and the enumeration resistance is thought through rather than asserted. That idea is buried in a document that also defines a resource type, a lifecycle state machine, a uniqueness contract, and a GDPR-adjacent residency profile, none of which need to move at the same speed. Split it, cut the indicator methods to the three that RFC 7644 already named, fix the four scimType pairings, and what's left is a short document with a clear argument and a charter question the chairs can actually answer.