Tightening up privacy in Matrix

Hi all,

A few weeks ago there was some discussion around the privacy of typical Matrix configurations, particularly how Riot's default config uses vector.im as an Identity Server (for discovering users on Matrix by their email address or phone number) and scalar.vector.im as an Integration Manager (i.e. the mechanism for adding hosted bots/bridges/widgets into rooms). This means that Riot, even if using a custom homeserver and running from a custom Riot deployment, will try to talk to *.vector.im (run by New Vector; the company formed by the core Matrix team in 2017) for some operations unless an alternative IS or IM has been specified in the config.

We haven't done as good a job at explaining this as we could have, and this blog post is a progress update on how we're fixing that and improving other privacy considerations in general.

Firstly, the reason Riot is configured like this is for the user's convenience: in general, we believe most users just want to discover other people on Matrix as easily as possible, and a logically-centralised server for looking up user matrix IDs by email/phone number (called third party IDs, or 3PIDs) is the only comprehensive way of doing so. Decentralising this data while protecting the privacy of the 3PIDs and their matrix IDs is a Hard Problem which we're unaware of anyone having solved yet. Alternatively, you could run a local identity server, but it will end up having to delegate to a centralised identity server anyway for IDs it has no other way to know about. Similarly, providing a default integration server that just works out of the box (rather than mandating the user configures their own) is a matter of trying to keep Riot's UX simple, especially when onboarding users, and especially given Riot's reputation for complexity at the best of times.

That said, the discussion highlighted some areas for improvement. Specifically:

1. When doing work on making Matrix GDPR compliant back in May 2018, we set up a single privacy policy for the services we run, and got users to agree to it by locking them out of the matrix.org homeserver until they did. However, we missed that users not on the matrix.org homeserver might still be using our Identity Service (IS) & Integration Manager (IM) without accepting the privacy policy. Over the last few weeks we've been working on addressing this - it turns out that it's a pain to fix, given the Identity Service doesn't have the concept of users, so tracking which users have agreed to the policy policy or not means some fairly major changes. The current proposal is to change the Identity Service to use a form of OpenID to authenticate users against their homeserver in order to check they've accepted the IS's terms of use - see MSC2140 for the gory details.

Meanwhile, Riot is being updated to prompt the user to accept the IS & IM terms of use (if different to the HS's), and thus make it crystal clear to the user that they are using an IS & IM and that they have the option not to if desired - see https://github.com/vector-im/riot-web/issues/10167 and associated issues. This includes also explicitly prompting the user as to whether they want 3PIDs they provide at registration to be discoverable, as per https://github.com/vector-im/riot-web/issues/10091.

2. Riot on iOS & Android gives the option of scanning your local addressbook to discover which of your contacts are on Matrix. The wording explaining this wasn't clear enough on Android - which we promptly fixed. Separately, the contact details sent to the server are currently not obfuscated. This is partially because we hadn't got to it, and partially because obfuscating them doesn't actually help much with privacy, given an attacker can just scan through possible obfuscated phone numbers and email addresses to deobfuscate them. However, we've been working through obfuscating the contact details anyway by hashing as per MSC2134, which has all the details. We're also adding an explicit lookup warning in Riot/Web, as per https://github.com/vector-im/riot-web/issues/10093.

3. There was a bug where Riot/Web was querying the Integration Manager every time you opened a room, even if that room had no integrations (actually, it did it 3 times in a row). This got fixed and released in Riot/Web 1.2.2 back on June 19th.

4. Matrix needs to authenticate whether events were actually sent by the server that claimed to send them. We do this by having servers sign their events when they create them, and publishing the public half of their signing keys for anyone to query. However, this poses problems if you receive an event which is signed by a server which isn't currently online. To solve this, we have the concept of trusted_key_servers (aka notary servers), which your server can query to see if they know about the missing server's keys. By default, matrix.org is configured as Synapse's trusted notary, but you can of course change this. If you choose an unreliable server as the notary (e.g. by not setting one at all) then there's a risk that you won't be able to look up signing keys, and a splitbrain will result where your server can't receive certain events, but other servers in the room can. This can then result in your server being unable to participate in the room entirely, if it's missing key events in the room's lifetime.

   Our plan here is to get rid of notaries entirely by changing how event signing works as per MSC1228, but this is going to take a while. Meanwhile we're going to check Synapse's code to ensure it doesn't talk to the notary server unnecessarily. (E.g. it should be caching the signing keys locally, and it should only use the notary server if the remote server is down.)

5. When doing VoIP in Matrix, clients need to use a TURN server to discover their network conditions and perform firewall traversal. The TURN server should be specified by your homeserver (and each homeserver deployment should ideally include a TURN server). However, for users who have not configured a TURN server, Riot (on all 3 platforms) defaulted to use Google's public STUN service (stun.l.google.com). STUN is a subset of TURN which provides firewall discovery, but not traffic relaying. This slightly increased the chances of calls working for users without a proper TURN server, but not by much - and rather than fall back to Google, we've decided to simply remove it from Riot (e.g. https://github.com/matrix-org/matrix-ios-sdk/commit/24832a2b14fb72ae6f051d5aba40262d11eef65d). This means that VoIP might get less reliable for users who were relying on this fallback, but you really should be running your own TURN server anyway if you want VoIP to work reliably on your homeserver.

6. We should make it clearer in Riot that device names are world-readable, and not just for the user's own personal reference. This is https://github.com/vector-im/riot-web/issues/10216

As you can see, much of the work on improving these issues is still in full swing, although some has already shipped. As should also be obvious, these issues are categorically not malicious: Matrix (and Riot) literally exists to give users full control and autonomy over their communication, and privacy is a key part of that. These are avoidable issues which can and will be solved. It's worth noting that we have to prioritise privacy issues alongside all the other development in Matrix however: there's no point in having excellent privacy if there are other bugs stopping the platform from being usable.

We'll do another blog post to confirm once most of the fixes here have landed - meanwhile, hopefully this post provides some useful visibility on how we're going about improving things.