MSC2778: Providing authentication method for appservice users

proposals/2778-appservice-login.md

@@ -0,0 +1,135 @@
+# MSC2778: Providing authentication method for appservice users
+
+Appservices within Matrix are increasingly attempting to support End-to-End Encryption. As such, they
+need a way to generate devices for their users so that they can participate in E2E rooms. In order to
+do so, this proposal suggests implementing an appservice extension to the
+[`POST /login` endpoint](https://matrix.org/docs/spec/client_server/r0.6.0#post-matrix-client-r0-login).
+
+Appservice users do not usually need to log in as they do not need their own access token, and do not
+traditionally need a "device". However, E2E encryption demands that at least one user in a room has a
+Matrix device which means bridge users need to be able to generate a device on demand. In the past,
+bridge developers have used the bridge bot's device for all bridge users in the room, but this causes
+problems should the bridge wish to only join ghosts to a room (e.g. for DMs).
+
+Another advantage this provides is that an appservice can now be used to generate access tokens for
+any user in its namespace without having to set a password for that user, which may be useful where
+maintaining password(s) in the configuration is undesirable.
+
+## Proposal
+
+A new `type` is to be added to `POST /login`: `m.login.application_service`
+
+The `/login` endpoint may now take an `access_token` in the same way that other
+authenticated endpoints do. No additional parameters should be specified in the request body.
+
+Example request
+
+```json
+{
+  "type": "m.login.application_service",
+  "identifier": {
+    "type": "m.id.user",
+    "user": "_bridge_alice"
+  }
+}
+```
+
+Note: Implementations MUST use the `identifier.type`=`m.id.user` method of specifying the
+localpart. The deprecated top-level `user` field **cannot** use this login flow type. This
+is deliberate so as to coax developers into using the new identifier format when implementing
+new flows.
+
+The response body should be unchanged from the existing `/login` specification.
+
+If one of the following conditions are true:
+
+- The access token is not provided
+- The access token does not correspond to an appservice
+- Or the user has not previously been registered
+
+Then the servers MUST reject with HTTP 403, with an `errcode` of `"M_FORBIDDEN"`.
+
+If the access token DOES correspond to an appservice but the user is not inside its namespace,
+then the `errcode` must be `"M_EXCLUSIVE"`.
+
+Homeservers should ignore the `access_token` parameter if a type other than
+`m.login.application_service` has been provided.
+
+Appservices creating **new** users can still use the `/register` endpoint to generate an `access_token` / `device_id`
+but for existing users, the `/login` endpoint can be used instead.
+
+## Potential issues
+
+This proposal means that there will be more calls to make when setting up a appservice user, when
+using encryption. While this could be done during the registration step, this would prohibit creating
+new devices should the appservice intentionally or inadvertently have lost the client-side device data.
+
+## Alternatives
+
+### 1. Include the token in the `/login` request body
+
+One minor tweak to the current proposal could be to include the token as part of the auth data, rather than
+being part of the header/params to the request. An argument could be made for either, but since the specification
+expects the appservice to pass the token this way in all requests, including `/register`, it seems wise to keep
+it that way.
+
+### 2. Use implementation specific "shared secret" authentication
+
+Some community members have used homeserver implementation details such as a "shared secret" authentication method to
+log into the accounts without having to use the /login process at all. Synapse provides such a function,
+but also means the appservice can now authenticate as any user on the homeserver. This is undesirable from a
+security standpoint.
+
+### 3. Keep using `/register` solely
+
+A third option could be to create a new endpoint that simply creates a new device for an appservice user on demand.
+Given the rest of the matrix eco-system does this with /login, and /login is already extensible with `type`, it would
+create more work for all parties involved for little benefit.
+
+Finally, `POST /register` does already return a `device_id` and `access_token` so appservices
+could store this information rather than calling `POST /login` at all. This does however present a few problems:
+
+- Quite a few appservices which only support unencrypted messaging do not use/store the `device_id`/`access_token` from a register call.
+  In the event that an appservice eventually gains the ability to support encryption, they would be unable to fetch a new `device_id`/
+  `access_token` for any existing users (as `/register` would fail for an existing user).
+- If user tokens were lost or exposed, there is no way to programattically create new access tokens for these users.
+- Finally, if a user was registered externally and the appservice would like to masquerade as it, it would be unable to fetch
+  an access token for that user.
+
+While `POST /register` does work, it is impactical as the sole method of fetching an access token.
+
+## Security considerations
+
+Appservices could use this new functionality to generate devices for any userId that are within its namespace e.g. setting the
+user namespace regex to `@.*:example.com` would allow appservice to control anyone on the homeserver. While this sounds scary, in practice
+this is not a problem because:
+
+- Appservice namespaces are maintained by the homeserver admin. If the namespace were to change, then it's reasonable
+  to assume that the server admin is aware. There is no defense mechanism to stop a malicious server admin from creating new
+  devices for a given user's account as they could also do so by simply modifying the database.
+
+- While an appservice *could* try to masquerade as a user maliciously without the server admin expecting it, it would still
+  be bound by the restrictions of the namespace. Server admins are expected to be aware of the implications of adding new
+  appservices to their server so the burden of responsibility lies with the server admin.
+
+- Appservices already can /sync as any user using the `as_token` and send any messages as any user in the namespace, the only
+  difference is that without a dedicated access token they are unable to receive device messages. While in theory this
+  does make them unable to see encrypted messages, this is not designed to be a security mechanism.
+
+In conclusion this MSC only automates the creation of new devices for users inside an AS namespace, which is something
+a server admin could already do. Appservices should always be treated with care and so with these facts in mind the MSC should
+be considered secure.
+
+## Unstable prefix
+
+Implementations should use `uk.half-shot.msc2778.login.application_service` for `type` given in the
+`POST /login` until this lands in a released version of the specification.
+
+## Implementations
+
+The proposal has been implemented by a homeserver, a bridge SDK and two bridges:
+
+- [synapse](https://github.com/matrix-org/synapse/pull/8320)
+- [mautrix-python](https://github.com/tulir/mautrix-python/commit/12d7c48ca7c15fd3ff61608369af1cf69e289aeb)
+- [mautrix-whatsapp](https://github.com/tulir/mautrix-whatsapp/commit/ead8a869c84d07fadc7cfcf3d522452c99faaa36)
+- [matrix-appservice-bridge](https://github.com/matrix-org/matrix-appservice-bridge/pull/231/files#diff-5e93f1b51d50a44fcf0ca46ea1793c1cR851-R864)