matrix-spec-proposals/proposals/2078-homeserver-password-resets.md

MSC2078 - Sending Third-Party Request Tokens via the Homeserver

This MSC proposes removing the current requirement of the identity server to send third-party request tokens, and allows homeservers to implement the functionality instead. These request tokens are used to verify the identity of the request author as an owner of the third-party ID (3PID). This can be used for binding a 3PID to an account, or for resetting passwords via email or SMS. The latter is what this proposal mainly focuses on, but be aware that it allows for any task that requires requesting a token for a 3PID to be taken on by the homeserver instead of the identity server.

The intention is to put less trust in the identity server, which is currently one of the most centralised components of Matrix. As it stands, an attacker in control of a identity server can reset a user's password if the identity server is considered trusted by that homeserver, and the user has registered at least one 3PID. This is due to the identity server handling the job of confirming the user's control of that identity.

The MSC seeks to clarify that homeservers can take on the responsibility of sending password reset tokens themselves, and a new response field that will aid homeservers in doing so.

Background

Currently when a client requests a 3PID token, it makes a call to one of the /requestToken endpoints on the homeserver. For instance, during password resets, a token is requested from either /_matrix/client/r0/account/password/email/requestToken or /_matrix/client/r0/account/password/msisdn/requestToken, depending on the medium of the 3PID. These requests are supplied all the necessary details as well as an id_server field containing the domain address of a identity server trusted by the homeserver.

In order to facilitate these requests, the homeserver will simply proxy them to the identity server. The IS will send out a token via email or sms, the user will click a link or enter the token into their client, and either the client or the user's browser will make a request directly to the identity server with the token for verification. The IS then informs the homeserver that verification was successful. At this point you can likely see that there is potential for abuse here, so instead Homeservers should be given the option to stop proxying the request to the identity server, and instead just send and validate the token themselves.

Proposal

The homeserver should be allowed to either proxy /requestToken requests or handle them itself. Specifically, this means that the homeserver can both send password reset tokens (via email or SMS), as well as accept requests on an arbitrary endpoint (with the same parameters as /_matrix/identity/api/v1/validate/email/submitToken) to verify that token.

One additional complication that in the case of SMS, just a code is sent to a person's phone. This is then given to the client, but the client may not know where to send the code now, as it doesn't know whether the homeserver or identity server generated it.

In order to combat this problem, the field submit_url MUST be added in the response from all of the variants of /requestToken in the Client-Server API, if and only if the verification message contains a code the user is expected to enter into the client (for instance in the case of a short code through SMS). This URL is simply where the client should submit this token. The endpoint should accept the same parameters as /_matrix/identity/api/v1/validate/{3pid_type}/submitToken in the Identity Service API. The only recommendation to homeserver developers for this endpoint's path is to not be exactly the same as that of the identity server, in order to prevent clashes between setups running both an identity server and homeserver on the same domain. If submit_url is omitted, the client MUST continue the same behaviour from before, which is to send the token to the identity server directly. This is intended for backwards compatibility with older servers.

If the client receives a response to /requestToken with submit_url, it MUST accept a token from user input, then make a POST request to the content of submit_url with the sid, client_secret and user-entered token. submit_url can lead to anywhere the homeserver deems necessary for verification. To be clear the content of id_server does not matter here, the client should just submit a POST request to the value of submit_url. Additionally data MUST be submitted as a JSON body.

An example exchange from the client's perspective is shown below:

POST https://homeserver.tld/_matrix/client/r0/account/password/email/requestToken

{
  "client_secret": "monkeys_are_AWESOME",
  "email": "alice@homeserver.tld",
  "send_attempt": 1,
  "id_server": "id.example.com"
}

If the server responds with a submit_url field, it means the client should collect a token from the user and then submit it to the provided URL.

{
  "sid": "123abc",
  "submit_url": "https://homeserver.tld/_homeserver/password_reset/msisdn/submitToken"
}

Since a submit_url was provided, the client will now collect a token from the user, say "123456", and then submit that as a POST request to the "submit_url".

POST https://homeserver.tld/_homeserver/password_reset/msisdn/submitToken

{
  "sid": "123abc",
  "client_secret": "monkeys_are_AWESOME",
  "token": "123456"
}

The client will then receive an appropriate response:

{
  "success": true
}

If the client did not receive a submit_url field, they should instead assume that verification will be completed out of band (e.g. the user clicks a link in their email and makes the submitToken request with their web browser).

Dismissed Alternatives

Consideration was taken not to make id_server an optional field. Let's assume for a moment that it was optional. Now, a client could send a request to /requestToken omitting the id_server field. The homeserver however has opted to continue proxying /requestToken to the identity server, even though it knows this is potentially insecure. The homeserver now has no idea which identity server to proxy the request to, and must return a failure to the client. The client could then make another request with an id_server, but we've now made two requests that ended up in the same outcome, instead of one, in hopes of saving a very small amount of bandwidth by omitting the field originally.

At some point we should look into removing the id_server field altogether and removing any email/SMS message sending from the identity server. This would drastically reduce the amount of trust needed in the identity server and its required ability. This is, however, a good first step.

Tradeoffs

If homeservers choose to not proxy the request, they will need to implement the ability to send emails and/or SMS messages. This is left as a detail for the homeserver implementation.