On why you should never trust an event.

proposals/2801-untrusted-event-data.md

@@ -0,0 +1,185 @@
+# MSC2801: Make it explicit that event bodies are untrusted data
+
+As the Matrix Specification stands today, it is easy for client developers to
+assume that all events will be structured as documented. For example, the
+`displayname` key of an [`m.room.member`
+event](https://matrix.org/docs/spec/client_server/r0.6.1#m-room-member) is
+specified to be "`string` or `null`"; and the `info` key of an [`m.image`
+message event](https://matrix.org/docs/spec/client_server/r0.6.1#m-image) must
+be a Javascript object.
+
+In reality, these are not safe assumptions. This MSC proposes that the
+specification be updated to make it clear that developers should treat all
+event data as untrusted.
+
+## Reasons why events may not match the specification
+
+Firstly, let's examine the reasons why such malformed events can exist. Some of
+these reasons may appear to have trivial solutions; these are discussed below.
+
+ 1. Most obviously, Synapse as it currently stands does very little validation
+    of event bodies sent over the Client-Server API. There is nothing stopping
+    any user sending a malformed `m.room.message` event with a simple `curl`
+    command or Element-Web's `/devtools`.
+
+    Any event sent in this way will be returned to the clients of all the other
+    users in the room.
+
+ 2. In order to allow the Matrix protocol to be extensible, server
+    implementations must tolerate unknown event types, and allow them to be
+    passed between clients. It is obvious that a pair of custom clients
+    implementing a `com.example.special.event` event type cannot rely on a
+    standard server implementation to do any validation for them.
+
+    However, this problem extends to event types and keys which have been added
+    to the specification. For example, the [`m.room.encrypted` event
+    type](https://matrix.org/docs/spec/client_server/r0.6.1#m-room-encrypted)
+    was added in Client-Server API r0.4.0. It therefore follows that a server
+    implementing CS-API r0.3.0 would have no way to validate an
+    `m.room.encrypted` event, so if a client is connected to such a server, it
+    could receive malformed events.
+
+ 3. To extend from point 2, the problem is not even resolved by upgrading the
+    server. There may now be rooms which contain historical events which would
+    no longer be accepted, but these will still be served by the server.
+
+    This problem also applies to non-room data such as account data. For
+    example, Client-Server API r0.6.0 added the [`m.identity_server` account
+    data event
+    type](https://matrix.org/docs/spec/client_server/r0.6.1#m-identity-server).
+    It is possible, if unlikely, that a client could have uploaded an
+    `m.identity_server` event before the administrator upgraded the server.
+
+ 3. Event redaction removes certain keys from an event. This is a bit of a
+    trivial case, though it is worth noting that the rules for event redaction
+    vary between room versions, so it's possible to see a variety of "partial"
+    events.
+
+ 4. All the cases above can occur without federation. Federation adds
+    additional complexities due to the structure of Matrix rooms. In
+    particular, a server implementation cannot simply ignore any malformed
+    events since such events may either be critical to the structure of the
+    room (for example, they may be `m.room.membership` events), or at the very
+    least ignoring them would leave "holes" in the event graph which would
+    prevent correct back-pagination.
+
+## Ideas for mitigating the problem
+
+The problems above appear to have some easy solutions. Let's describe some of
+those ideas, before considering the fundamental problem that none of them can
+solve.
+
+### Validate all events when they are submitted
+
+In this idea, we would require all server implementations to strictly validate
+any data which is sent over the Client-Server API, to ensure that it complied
+with the specified formats.
+
+This evidently solves problem 1 above, in that it would prevent local users from
+creating malformed events of any event types that the server supports; however,
+it would do nothing to address any of the other problems.
+
+### Validate events at the point they are served to a client
+
+We could require that server implementations validate any data that they are
+about to serve to a client; we might recommend that malformed data be stripped
+out of the response, or redacted, or similar.
+
+It is worth mentioning that this would be tricky to implement efficiently on
+the server side, but it at least helps address most of the problems above, such
+as historical data, or malformed events received over federation.
+
+### Have servers re-validate data on upgrade
+
+Similar to the previous idea, but rather than validating data each time it is
+served to a client, any stored data could be re-validated to check that it
+complies with new validation requirements.
+
+This could be more efficient in the case that changes to validation rules are
+rare, but it could still be a huge amount of data processing on a large server.
+
+### Create new room versions which require events to be well-formed
+
+As outlined above, one of the big problems in this area is how we deal with
+events sent over federation; in particular, if subsets of the servers in a room
+have different ideas as to which events are "valid", then their concepts of the
+room state can begin to drift, and the room can eventually become
+"split-brained". This makes it hard to simply say, for example,
+"`m.room.member` events with a non-string `displayname` are invald and should
+not form part of the room state": we have a risk that some servers will accept
+the event, and some will not.
+
+One approach to solving this is via [room
+versions](https://matrix.org/docs/spec/index#room-versions). By specifying that
+a change of rules only applies for a future room version, we can eliminate this
+potential disagreement.
+
+The process of changing a room from one version to another is intrusive, not
+least because it requires that all servers in a room support the new room
+version (or risk being locked out). For that reason, it is extremely
+undesirable that any new feature require a new room version: whenever possible,
+it should be possible to use new features in existing rooms. It therefore
+follows that we cannot rely on room versions to provide validation of event
+data.
+
+### Create a single new room version which applies all event validation rules
+
+This idea is included for completeness, though it is unclear how it would work
+in practice.
+
+It has been suggested that we create a new room version which explicitly states
+that events which fail the current event schema, whatever that is at that
+moment in time, should be rejected.
+
+Let's imagine that in future, the `m.room.member` event schema is extended to
+include an optional `location` key, which, if given, must be a string. The
+implication of this idea is that servers should reject any `m.room.member`
+event whose `location` is not a string. We now have a problem: any servers in
+the room which are updated to the latest spec will reject such malformed
+events, but any other servers yet to be upgraded will allow it. So is that user
+in the room or not?
+
+Even if all the servers in the room can be upgraded at once, what about any
+`m.room.member` events which were sent before the rule change?
+
+## The fundamental problem
+
+The ideas above all mitigate the problems discussed earlier to a greater or
+lesser extent, and may indeed be worth doing on their own merits. However, none
+of them can address the problem of outdated server implementations.
+
+For example, consider the case of a new key being added to an event body, say
+`m.relates_to`. Now, we may have decided as above that all new specced keys
+must be validated by the server, so `vN+1` of Synapse dutifully implements such
+validation and refuses to accept events with a malformed `m.relates_to`.
+
+The problem comes for users whose server is still Synapse `vN`. It knows
+nothing of `m.relates_to`, so accepts and passes it through even if
+malformed. The only potential solution is for clients seeking to implement
+`m.relates_to` to refuse to talk to servers which do not declare support for
+it.
+
+However, this is an entirely client-side feature: it is illogical to require
+that servers must be upgraded before it can be used. Consider that the hosted
+element-web at `https://app.element.io` is upgraded to support the new feature;
+in this scenario, that would lock out any user whose homeserver had not yet
+been upgraded. This is not an acceptable user experience.
+
+In short, we are left with the reality that clients must still handle the
+unvalidated data.
+
+## Conclusions
+
+Short of closely coupling server and client versions - which violates the
+fundamental ethos of the Matrix project - there is nothing that can completely
+prevent clients from having to handle untrusted data.
+
+With that in mind, the advantages of the ideas above are diminished. If clients
+must handle untrusted data in some circumstances, why not in all? "You can
+trust the content of this data structure, provided you have checked that the
+server knows how to validate it, in which case you need to treat it as
+untrusted" is not a useful message for a client developer.
+
+It may be possible to assert that specific, known cases can be treated as
+trusted data, but these should be called out as specific cases. The default
+message should be that clients must treat all event data as untrusted.