The actual problem isn't the network connection, it's the untrusted data. There's nothing you can do about that if your data is intentionally coming from untrus

The actual problem isn't the network connection, it's the untrusted data. There's nothing you can do about that if your data is intentionally coming from untrusted sources, which is absolutely the case for a mail server. Of course you can reduce attack surfaces by having things be tunneled and/or proxied, to have close-to-minimal network access to some appliance, but that won't fully save you from this sort of problem. In my opinion, the only obvious way out is to go category-by-category and eliminate each type of flaw by-design. I think even in an ideal world you could never accomplish this 100% of the way, or at least we're not close to a world where you could right now, but you can combine this with a layered approach to security, trying to eliminate single points of failure, adding hardening anywhere you can, and making tampering both as apparent and as unreproducible as possible (i.e.: ASLR harder and more often, randomize the order of things, don't expose internal IDs, etc.)