Global Token Revocation

2.1. Terminology

This specification uses the terms "Access Token", "Authorization Code", "Authorization Endpoint", "Authorization Server" (AS), "Client", "Client Authentication", "Client Identifier", "Client Secret", "End-User", "Grant Type", "Protected Resource", "Redirection URI", "Refresh Token", "Resource Owner", "Resource Server" (RS) and "Token Endpoint" defined by [RFC6749], and the terms "OpenID Provider" (OP) and "ID Token" defined by [OpenID].¶

This specification uses the term "Identity Provider" (IdP) to refer to the Authorization Server or OpenID Provider that is used for End-User authentication.¶

TODO: Replace RFC6749 references with OAuth 2.1¶

2.2. Roles

In a typical OAuth deployment, the OAuth client obtains tokens from the authorization server when a user logs in and authorizes the client. In many cases, the method by which a user logs in at the authorization server is through an external identity provider.¶

For example, a mobile chat application is an OAuth Client, and obtains tokens from its backend server which stores the chat messages. The mobile chat backend plays the OAuth roles of "Resource Server" and "Authorization Server".¶

In some cases, the user will log in to the Authorization Server using an external (e.g. enterprise) Identity Provider. In that case, when a user logs in to the chat application, the backend server may play the role of an OAuth client (or OpenID or SAML relying party) to the Identity Provider in a new authorization or authentication flow.¶

3.1. Revocation Endpoint

The Global Token Revocation endpoint is a URL at the authorization server which accepts HTTP POST requests with parameters in the HTTP request message body using the application/json format. The Global Token Revocation endpoint URL MUST use the https scheme.¶

If the authorization server supports OAuth Server Metadata ([RFC8414]), the authorization server SHOULD include the URL of their Global Token Revocation endpoint in their authorization server metadata document using the global_token_revocation_endpoint parameter as defined in Section 5.¶

The authorization server MAY alternatively register the endpoint with tools that will use it.¶

3.2. Revocation Request

The request is a POST request with an application/json body containing a single property subject, the value of which is a Security Event Token Subject Identifier as defined in "Subject Identifiers for Security Event Tokens" [RFC9493].¶

In practice, this means the value of subject is a JSON object with a property format, and at least one additional property depending on the value of format.¶

The request MUST also be authenticated, the particular authentication method and means by which the authentication is established is out of scope of this specification, but may include OAuth 2.0 Bearer Token [RFC6750] or a JWT [RFC7523].¶

The following example requests that all tokens for a user identified by an email address be revoked:¶

POST /global-token-revocation
Host: example.com
Content-Type: application/json
Authorization: Bearer f5641763544a7b24b08e4f74045

{
  "sub_id": {
    "format": "email",
    "email": "user@example.com"
  }
}

If the user identifier at the authorization server is known by the system making the revocation request, the request can use the "Opaque Identifer" format to provide the user identifier:¶

POST /global-token-revocation
Host: example.com
Content-Type: application/json
Authorization: Bearer f5641763544a7b24b08e4f74045

{
  "sub_id": {
    "format": "opaque",
    "id": "e193177dfdc52e3dd03f78c"
  }
}

If it is expected that the authorization server knows about the user identifier at the IdP, the request can use the "Issuer and Subject Identifier" format:¶

POST /global-token-revocation
Host: example.com
Content-Type: application/json
Authorization: Bearer f5641763544a7b24b08e4f74045

{
  "sub_id": {
    "format": "iss_sub",
    "iss": "https://issuer.example.com/",
    "sub": "af19c476f1dc4470fa3d0d9a25"
  }
}

3.3. Revocation Expectations

Upon receiving a revocation request, authorizing the request, and validating the identified user, the Authorization Server:¶

• MUST revoke all active refresh tokens¶

• SHOULD invalidate all access tokens, although it is recognized that it might not be technically feasible to invalidate access tokens (see Section 4 below)¶

• MUST re-authenticate the user before issuing new access tokens or refresh tokens¶

3.4. Revocation Response

This specification indicates success and error conditions by using HTTP response codes, and does not define the response body format or content.¶

6.1. Authentication of Revocation Request

While Section 3.2 requires that the revocation request is an authenticated request, the specifics of the authentication are out of scope of this specification.¶

Since the revocation request ultimately has wide-reaching effects (a user is expected to be logged out of all devices), this presents a new Denial of Service attack vector. As such, the authentication used for this request SHOULD be narrowly scoped to avoid granting unnecessary privileges to the caller.¶

For example, if using OAuth Bearer Tokens, the token SHOULD be issued with a single scope that enables it to perform the revocation request, and no other type of token issued should include this scope.¶

If the authorization server is multi-tenant (supports multiple customers) through different identity providers, each identity provider SHOULD use its own scoped credential that is only authorized to revoke tokens for users within the same tenant.¶

6.2. Enumeration of User Accounts

Typically, an API that accepts a user identifier and returns different statuses depending on whether the user exists would provide an attack vector allowing enumeration of user accounts. This specification does require a "User Not Found" response, so would normally fall under this category. However, requests to the endpoint defined by this specification are required to be authenticated, so this is not considered a public endpoint.¶

If the tool making the request is compromised, and the attacker can impersonate the requests from this tool (either by coercing the tool to make the request, or by extracting the credentials), then the attacker would be able to enumerate user accounts. However, since the request is not just testing the presence of a user account, but actually revoking the tokens associated with the user if successful, this would likely be easily visible in any audit logs as many users' tokens would be revoked in a short period of time.¶

To mitigate some of the concerns of providing such a powerful API endpoint, the users that a particular client can request revocation for SHOULD be limited, and the authentication of the request SHOULD be used to scope the possible user revocation list to only users authorized to the client as described in Section 6.1.¶

For example, a multi-tenant identity provider that uses different signing keys for users associated with different tenants, can also use the same signing keys to authenticate revocation requests, such as creating a JWT to use as client authentication as described in [RFC7523]. This enables the authorization server receiving the request to only accept revocation requests for users that are associated with the particular tenant at the identity provider.¶

6.3. Malicious Authorization Server

From the point of view of an identity provider that supports integrations with multiple downstream applications, there is an opportunity for a downstream application to maliciously set up a Global Token Revocation endpoint to harvest user identifiers and authentication of the revocation requests.¶

Similarly as described in Section 6.1 above, each integration SHOULD be using separate authentication credentials, and each credential SHOULD be scoped as narrowly as possible, such that a malicious server that receives this authentication cannot replay it anywhere else to perform any actions on other systems.¶

7.1. OAuth Authorization Server Metadata

IANA has (TBD) registered the following values in the IANA "OAuth Authorization Server Metadata" registry of [IANA.oauth-parameters] established by [RFC8414].¶

Metadata Name: global_token_revocation_endpoint¶

Metadata Description: URL of the authorization server's global token revocation endpoint.¶

Change Controller: IESG¶

Specification Document: Section X of [[ this specification ]]¶

Metadata Name: global_token_revocation_endpoint_auth_methods_supported¶

Metadata Description: OPTIONAL. Indicates the list of client authentication methods supported by this endpoint.¶

Change Controller: IESG¶

Specification Document: Section X of [[ this specification ]]¶

A.1. RFC7009: Token Revocation

OAuth 2.0 Token Revocation [RFC7009] defines an endpoint for authorization servers that an OAuth client can use to notify the authorization server that a previously-obtained access or refresh token is no longer needed.¶

The request is made by the OAuth client. The input to the Token Revocation request is the token itself, as well as the client's own authentication credentials.¶

This differs from the Global Token Revocation endpoint which does not take a token as an input, but instead takes a user identifier as input. It is not called by OAuth clients, but is instead called by an external party such as a security monitoring tool or an identity provider that the user used to authenticate at the authorization server.¶

A.2. OpenID Connect Front-Channel Logout

OpenID Connect Front-Channel Logout provides a mechanism for an OpenID Provider to log users out of Relying Parties by redirecting the user agent.¶

While the logout request is the same direction as this draft describes, this relies on the redirection of the user agent, so is only applicable when the user is actively interacting with the application in a web browser.¶

The Global Token Revocation request works regardless of whether the user is actively using the application, and is also applicable to non-web based applications.¶

A.3. OpenID Connect Back-Channel Logout

OpenID Connect Back-Channel Logout provides a mechanism for an OpenID Provider to log users out of a Relying Party by making a back-channel POST request containing the user identifier of the user to log out.¶

This is the most similar existing logout specification to Global Token Revocation. However, there are still a few key differences that make it insufficient for the use cases enabled by Global Token Revocation.¶

OpenID Connect Back-Channel Logout requires Relying Parties to clear state of any sessions for the user, but doesn't mention anything about access tokens. It also says that refresh tokens issued with the offline_access scope "SHOULD NOT be revoked". This is a concretely different outcome than is described by Global Token Revocation, which requires the revocation of all refresh tokens for the user regardless of whether the refresh token was issued with the offline_access scope.¶

Additionally, OpenID Connect Back-Channel Logout assumes that the Relying Party implements OpenID Connect, which creates implementation challenges to use it when the Relying Party actually integrates with the identity provider using other specifications such as SAML.¶

Global Token Revocation works regardless of the protocol that the user uses to authenticate, so works equally well with OpenID Connect and SAML integrations.¶

A.4. Shared Signals Framework

The Shared Signals Framework at the OpenID Foundation provides two specifications that have functionality related to session and token revocation.¶

Continuous Access Evaluation Profile (CAEP) defines several event types that can be sent between cooperating parties. In particular, the "Session Revoked" event can be sent from an identity provider to an authorization server when the user's session at the identity provider was revoked. The main difference between this and the Global Token Revocation request is that the CAEP event is a signal that may or may not be acted upon by the receiver, whereas the Global Token Revocation request is a command that has a defined list of expected outcomes.¶

Risk Incident Sharing and Coordination (RISC) defines events that have somewhat stronger defined meanings compared to CAEP. In particular, the "Account Disabled" event has clear meaning and strongly implies that a receiver should also disable the specified account. However, RISC also has a mechanism for a user to opt out of sending events for their account, so it does not provide the same level of assurance as a Global Token Revocation request.¶

Lastly, it is more complex to set up a receiver for CAEP and RISC events compared to a receiver for the Global Token Revocation request, so if the receiver is only interested in supporting the revocation use cases, it is much simpler to support the single POST request described in this draft.¶