Mail, SSL and Postfix

In my previous post, I delineated a few reasons why I care about SSL for my blog and the xine bugzilla. What I did not talk about was about the email infrastructure for both. The reason is that, according to the very same threat model that I delineated in that post, it’s not as important for me to secure that part of the service.

That does not mean, though, that I never considered it, just I considered it not important enough yet. But after that post I realized it’s time to fix that hole and I’ve now started working on securing the email coming out of my servers as well as that coming through the xine server. But before going into details about that, let’s see why I was not as eager to secure the mail servers compared to the low-hanging fruit of my blog.

As I said in the previous post, what you have to identify two details: what information is important to defend, and who the attackers would be. In the case of the blog, as I said, the information was the email addresses of the commenters, and the attackers the other users of open, unencrypted wifi networks in use. In the case of email, the attackers in particular change drastically; the only people in a position to get access to the connections’ streams are the people in the hosting and datacenter companies, and if they made mistakes, your neighbours in the same datacenter. So for sure it’s not the very easy prey of the couple sitting at the Starbucks next to you.

The content, well, is a bit more interesting information. We all know that there is no real way to make email completely opaque to service providers unless we use end to end encryption such as GnuPG, so if you really don’t want your server admin to ever be able to tell what’s in your email, that’s what you should do. But even then, there is something that (minus protocol-level encryption) is transmitted in cleartext: the headers, the so-called metadata, that stirred the press so much last year. So once again it’s the address of the people you contact that could be easily leaked, even with everything else being encrypted. In the case of xine, the mail server handles mostly bugzilla messaging, and it is well possible that it sends over, without encryption, the comments on security bugs, so reducing the risk of that information leaking is still a good idea.

Caveat emptor in all of this post, though! In the case of the xine mail server, the server handles both inbound and outbound messages, but at the same time it does not ever let users access their mailbox; the server itself is a mail router, rather than a full mail service. This is important, becuase otherwise I wouldn’t be able to justify my sloppiness on covering SSL support for the mail! If your server hosts mailboxes or allows direct mail submission (relay), you most definitely need to support SSL as then it’s a client-server connection which is attackable by the Starbucks example above.

So what needs to be done to implement this? Well, first you need to remember that a mail router like the one I described above requires SSL in two directions: when it receive a message it should be able to offer SSL to the connecting client, and when it sends a message it has to request SSL to the remote server too. In a perfect set up, the client also offers a certificate to prove who it is. This means that you need a certificate that works both as a server and as a client certificate; thankfully, StartSSL supports that for Class 2 certificates, even if they are named for web servers, they work just fine for mail servers too.

Unfortunately, the same caveat that apply to HTTP certificates, apply to mail servers: ciphers and protocol versions combinations. Unfortunately, while Qualys has SSL Labs to qualify the SSL support in your website, I know of no similar service for mail routers, and coming up with one is not trivial, as you would want to make sure not to become a relay-spammer by mistake, and the only way to judge the message pushing of the server is to trick it into sending a message to your own service back, which should not be possible on a properly non-open relay of a server.

So the good news is that of all of the xine developers with an alias on the domain have a secure server when routing mail to them, so the work I’ve been done is not for nothing. The other note is that a good chunk of the other users in Bugzilla uses GMail or similar big hosting providers. And unlike others I actually find this a good thing, as it’s much more likely that the lonely admin of a personal mail server (like me for xine) would screw up encrypion, compared to my colleagues over at GMail. But I digress.

The bad news is that not only there is no way to assess the quality of the configuration of a mail server, but at least for the case of Postfix you have only a ternary setting about TLS: yes always, yes if client requests it (or if the server provides the option, when submitting mail), or not at all. There is no way to set up policy so that e.g. gmail servers don’t get spoofed and tricked into sending the messages over a clear text connection. A second bad news is that I have not been able to get Postfix to validate the certificates either as server or client, likely caused by the use of opportunistic TLS rather than forcing TLS support. And the last problem is that servers connecting to submit mail will not fallback to cleartext if the TLS can’t be negotiated (either because of cipher or protocols), and will instead keep trying to negotiate TLS the same way.

Anyway, my current configuration for this is:

smtpd_tls_cert_file = /etc/ssl/postfix/server.crt
smtpd_tls_key_file = /etc/ssl/postfix/server.key
smtpd_tls_received_header = yes
smtpd_tls_loglevel = 1
smtpd_tls_security_level = may
smtpd_tls_ask_ccert = yes
smtpd_tls_mandatory_protocols = !SSLv2, !SSLv3

smtp_tls_cert_file = /etc/ssl/postfix/server.crt
smtp_tls_key_file = /etc/ssl/postfix/server.key
smtp_tls_security_level = may
smtp_tls_loglevel = 1
smtp_tls_protocols = !SSLv2, !SSLv3

If you have any suggestions on how to make this more reliable and secure, please do!