matrix-spec/proposals/1711-x509-for-federation.md

MSC1711: X.509 certificate verification for federation connections

TLS connections for server-to-server communication currently rely on an approach borrowed from the Perspectives project to provide certificate verification, rather than the more normal model using certificates signed by trusted Certificate Authorities. This document sets out the reasons that this has not been a success, and suggests that we should instead revert to the CA model.

Background: the failure of the Perspectives approach

The Perspectives approach replaces the conventional hierarchy of trust provided by the Certificate Authority model with a large number of "notary" servers distributed around the world. The intention is that the notary servers regularly monitor remote servers and observe the certificates they present; when making a connection to a new site, a client can correlate the certificate it presents with that seen by the notary servers. In theory this makes it very hard to mount a Man-in-the-Middle (MitM) attack, because it would require intercepting traffic between the target server and a large number of the notary servers.

It is notable that the Perspectives project itself appears to have largely been abandoned: its website has largely been repurposed, the Firefox extension does not work with modern versions of Firefox, the mailing list is inactive, and several of the (ten) published notary servers are no longer functional. The reasons for this are not entirely clear, though clearly it never gained widespread adoption.

When Matrix was originally designed in 2014, the Perspectives project was heavily active, and avoiding dependencies on the relatively centralised Certificate Authorities was attractive, in accordance with Matrix's design as a decentralised protocol. However, this has not been a success in practice.

Matrix was unable to make use of the existing notary servers (largely because we wanted to extend the protocol to include signing keys): the intention was that, as the Matrix ecosystem grew, public Matrix servers would act as notary servers. However, in practice we have ended up in a situation where almost ¹ every Matrix homeserver either uses matrix.org as the sole notary, or does no certificate verification at all. Far from avoiding the centralisation of the Certificate Authorities, the entire protocol is therefore dependent on a single point of control at matrix.org - and because matrix.org only monitors from a single location, the protection against MitM attacks is weak.

It is also clear that the Perspectives approach is poorly-understood. It is a common error for homeservers to be deployed behind reverse-proxies which make the Perspectives-based approach unreliable. The CA model, for all its flaws, is at least commonly used, which makes it easier for administrators to deploy (secure) homeservers, and allows server implementations to leverage existing libraries.

Proposal

We propose that Matrix homeservers should be required to present valid TLS certificates, signed by a known Certificate Authority, on their federation port.

In order to ease transition and give administrators time to switch to a signed certificate, we will continue to follow the current, perspectives-based approach for servers whose TLS certificates fail validation.

However, this fallback will be strictly time-limited, and Matrix S2S spec r0 will not accept self-signed certificates, nor will it include the tls_fingerprints property of the /_matrix/key/v2 endpoints. Synapse 1.0 will not accept self-signed certificates by default.

The matrix.org team will proactively attempt to reach out to homeserver administrators who do not update their certificates in the coming weeks.

The process of determining which CAs are trusted to sign certificates would be implementation-specific, though it should almost certainly make use of existing operating-system support for maintaining such lists. It might also be useful if administrators could override this list, for the purpose of setting up a private federation using their own CA.

It would also be useful for administrators to be able to disable the certificate checks for a whitelist of domains/netmasks. This would be useful for .onion domains (where a certificate is hard to obtain, and where server verification is provided at the network level), as well as for testing with IP literals.

Interaction with SRV records

With the use of SRV records, it is possible for the hostname of a homeserver to be quite different from the matrix domain it is hosting. For example, if there were an SRV record at _matrix._tcp.matrix.org which pointed to server.example.com, then any federation requests for matrix.org would be routed to server.example.com. The question arises as to which certificate server.example.com should present.

In short: the server should present a certificate for the matrix domain (matrix.org in the above example). This ensures that traffic cannot be intercepted by a MitM who can control the DNS response for the SRV record (perhaps via cache-poisoning or falsifying DNS responses).

This will be in line with the current requirements in the Federation API specification for the Host, and by implication, the TLS Server Name Indication ². It is also consistent with the recommendations of RFC6125 and the conventions established by the XMPP protocol (per RFC6120.

Interaction with .well-known files

MSC1708 proposes an alternative to SRV records, in the form of .well-known files. In this instance, a file at https://matrix.org/.well-known/matrix/server might direct requests to server.example.com.

In this case, server.example.com would be required to present a valid certificate for server.example.com.

Because the request for the .well-known file takes place over a validated TLS connection, this is not subject to the same DNS-based attacks as the SRV record, and this mechanism allows the owners of a domain to delegate responsibility for running their Matrix homeserver without having to hand over TLS keys for the whole domain.

Extensions

HTTP-Based Public Key Pinning (HPKP) and Certificate transparency are both HTTP extensions which attempt to work around some of the deficiencies in the CA model, by making it more obvious if a CA has issued a certificate incorrectly.

HPKP has not been particularly successful, and is deprecated in Google Chrome as of April 2018. Certificate transparency, however, is seeing widespread adoption from Certificate Authories and HTTP clients.

This proposal sees both technologies as optional techniques which could be provided by homeserver implementations. We encourage but do not mandate the use of Certificate Transparency.

Related work

The Perspectives approach is also currently used for exchanging the keys that are used by homeservers to sign Matrix events and federation requests (the "signing keys"). Problems similar to those covered here also apply to that mechanism. This is discussed at #1685.

Alternatives

There are well-known problems with the CA model, including a number of widely-published incidents in which CAs have issued certificates incorrectly. It is therefore important to consider alternatives to the CA model.

Improving support for the Perspectives model

In principle, we could double-down on the Perspectives approach, and make an effort to get servers other than matrix.org used as notary servers. However, there remain significant problems with such an approach:

• Perspectives remain complex to configure correctly. Ideally, administrators need to make conscious choices about which notaries to trust, which is hard to do, especially for newcomers to the ecosystem. (In practice, people use the out-of-the-box configuration, which is why everyone just uses matrix.org today).

• A correct implementation of Perspectives really needs to take into account more than the latest state seen by the notary servers: some level of history should be taken into account too.

Essentially, whilst we still believe the Perspectives approach has some merit, we believe it needs further research before it can be relied upon. We believe that the resources of the Matrix ecosystem are better spent elsewhere.

DANE

DNS-Based Authentication of Named Entities (DANE) can be used as an alternative to the CA model. (It is arguably more appropriately used together with the CA model.)

It is not obvious to the author of this proposal that DANE provides any material advantages over the CA model. In particular it replaces the centralised trust of the CAs with the centralised trust of the DNS registries.

Potential issues

Beyond the problems already discussed with the CA model, requiring signed certificates comes with a number of downsides.

More difficult setup

Configuring a working, federating homeserver is a process fraught with pitfalls. This proposal adds the requirement to obtain a signed certificate to that process. Even with modern intiatives such as Let's Encrypt, this is another procedure requiring manual intervention across several moving parts³.

On the other hand: obtaining an SSL certificate should be a familiar process to anybody capable of hosting a production homeserver (indeed, they should probably already have one for the client port). This change also opens the possibility of putting the federation port behind a reverse-proxy without the need for additional configuration. Hopefully making the certificate usage more conventional will offset the overhead of setting up a certificate.

Furthermore, homeserver implementations could provide an implementation of the ACME protocol and integration with Let's Encrypt, to make it easier for administrators to get started. (This would of course be implementation-specific, and administrators who wanted to keep control of the certificate creation process would be free to do so).

Inferior support for IP literals

Whilst it is possible to obtain an SSL cert which is valid for a literal IP address, this typically requires purchase of a premium certificate; in particular, Let's Encrypt will not issue certificates for IP literals. This may make it impractical to run a homeserver which uses an IP literal, rather than a DNS name, as its server_name.

It has long been the view of the matrix.org administrators that IP literals are only really suitable for internal testing. Those who wish to use them for that purpose could either disable certificate checks inside their network, or use their own CA to issue certificates.

Inferior support for hidden services (.onion addresses)

It is currently possible to correctly route traffic to a homeserver on a .onion domain, provided any remote servers which may need to reach that server are configured to route to such addresses via the Tor network. However, it can be difficult to get a certificate for a .onion domain (again, Let's Encrypt do not support them).

The reasons for requiring a signed certificate (or indeed, for using TLS at all) are weakened when traffic is routed via the Tor network. Administrators using the Tor network could disable certificate checks for .onion addresses.

Conclusion

We believe that requiring homeservers to present an X.509 certificate signed by a recognised Certificate Authority will improve security, reduce centralisation, and eliminate some common deployment pitfalls.

[1] It's possible to set up homeservers to use servers other than matrix.org as notaries, but only a minority are actually set up this way. ↩

[2] I've not been able to find an authoritative source on this, but most reverse-proxies will reject requests where the SNI and Host headers do not match. ↩

[3] Let's Encrypt will issue ACME challenges via port 80 or DNS (for the http-01 or dns-01 challenge types respectively). It is unlikely that a homeserver implementation would be able to control either port 80 or DNS responses, so we will be unable to automate a Let's Encrypt certificate request. ↩