STARTTLS Considered Harmful
agwa.name
agwa.name
That's actually a pretty good point and I wish more services used SRV records to determine which host and port to use. HTTP 2.0 was a chance to implement it for HTTP but the effort fizzled out after just a handful of comments in 2007 or so.
I also like the idea of squeezing other information into DNS, like certificates, gpg keys, why not HSTS-like information too? Say, a TXT record enforcing ssl, no exceptions (ssl=all). But all this hinges on DNSSEC which is very unpopular here.
Of course what we get instead is unspecified and even conflicting* behaviour, which I think just stinks. Many other standards have managed to get it together by specifying the use of SRV. I gave up trying to change the entrenched positions.
* there is language in the draft that advises against multiple connections, conflicting with the current "Happy Eyeballs" dual-stack resolution workaround.
CAs are a problem that DNSSEC doesn't solve, but DNSSEC solves tons of other problems without making the CA problem harder.
Centralized, hierarchical PKI is ok here, because it is better than what we have now (600 CAs for TLS, unsigned DNS, etc.) and nobody has a proposal for a fully decentralized system that ordinary people can understand (as demonstrated by GPG).
If you are starting a new protocol for a wide audience you cannot just say "encryption is required". Which encryption protocol will you require? And will you stick with it forever? Even when it will be broken?
So you want a bit of forward compatibility and you add a version identifier to wire protocol the or some other form of negotiation. Well done, now you are opening your self to protocol downgrade attacks. You can try to fight this as much as you want, but will always be back to square one: support for non-updated applications: drop them or support them even thou you know there is a security problem in supporting them?
All the non-public SMTP servers I have used in the last years require the use of STARTTLS-enabled connections. So they are dropping old clients, but they can do that because they have a tight control over the "environment" (their students, their employers, etc). Others do not have this luxury.
Yes, the crypto engineering can be bewildering for non-experts, but it's just engineering at the end of the day.
In many cases there are defenses against downgrade attacks. There is even hope for the web browsery scenario (have to be back compatible with SSL + connect to unknown/untrusted server, same as email): See http://tools.ietf.org/html/draft-bmoeller-tls-downgrade-scsv...
What should be the response of a client when a server tells it during the cipher negotiation that the only cipher that it (the server) can use are those considered insecure by client?
Should the client proceed for the sake of interoperability? Or should the client raise and error and say "sorry, no secure connection can be established"?
Whatever you answer to that question is, why doesn't it apply to STARTTLS As well? By not allowing STARTTLS, the server has negotiated a set of cipher (i.e. {NULL-ENC}) that the client considers insecure. What should the client do? Close the connection? (You can configure your program to do so right now.) Or continue because you really want to send that email? (This is, sadly, the current configuration of many applications.)
Back to the original point, I think that the important thing here is not only to embed secure layers in protocols, but also (and mainly) to convince the users that it is OK for an application to answer "Sorry, no secure connection is available. Yes, I could set up an insecure connection, but I will not do that."
Should the client proceed for the sake of interoperability? Or should the client raise and error and say "sorry, no secure connection can be established"?
These are solved problems though; it's not as though this is a new consideration. This has been a 'problem' since SSL's implementation. If you can't agree on a cipher spec, you don't connect.
The issue with your argument in particular is that you're basically saying 'we shouldn't require encryption because it's too complicated to make a few choices', which implies to me that you think not using encryption is a better choice (since any TLS implementation requires these issues). Your argument, then, is an argument against SSL/TLS encryption in any respect, as opposed to mandatory encryption.
Back to the original point, I think that the important thing here is not only to embed secure layers in protocols, but also (and mainly) to convince the users that it is OK for an application to answer "Sorry, no secure connection is available. Yes, I could set up an insecure connection, but I will not do that."
This is pretty common; in fact, most software will just say that it can't connect, with no more detailed explanation unless you dig into it. I've also had issues where the default configuration for postfix on Ubuntu (STARTLS enabled, self-signed certificate) is rejected by some clients, requiring me to disable authentication on those clients (which required a secure connection) and whitelist their IP addresses (insecure!).
If you have authentication in the first version of your protocol, you can survive downgrade attacks by authenticating the version number.
For example, my email server will accept any email encrypted or not, but will only allow authentication over a channel if it's encrypted. But encryption is not optional for the latter case.
Email clients that attempt to authenticate won't send passwords over unencrypted plain-text connections, so there's no real vulnerability, unless a client is broken.
It prevents casual attackers watching the wire i.e. network taps but doesn't prevent MITM at all as you can just proxy it with another self signed cert and it goes through.
I know this because our SMTP gateway does exactly that, transparently with postfix.
The downgrade problem isn't moot with other protocols (like IMAP or client-to-server SMTP) which are virtually always expected to be secure. And the other points, about how STARTTLS adds complexity and prevents protocol-agnostic termination, still apply.
It's not a protocol but an implementation problem.
This is not a protocol vulnerability though. Poorly written programs cause problems with any protocol.
That was the whole point of STARTTLS - to allow a way to start a tunnel but be backwards compatible to older clients.
The real problem was that they didn't account for MITM attacks.
https://www.fastmail.fm/help/technical/ssltlsstarttls.html
All the downgrade and pre-encryption MITM and complexity arguments are definitely worth repeating though. Straight up SSL is significantly better.
Point is, I think it's naive and unwise to assume that the crypto upgrade/wrapper is and always shall be TLS and then to build this assumption into standards in a manner that isn't future compatible. We are due a great many innovations in end-to-end crypto over the next few decades. For example, I'm hoping that new and interesting uses will be found for DANE. Maybe we can even bring back network-layer opportunistic encryption.
Worse still, most mail servers and programs have no checks whatsoever regarding certificates of other mail servers or clients. As long as it is able to crypt and decrypt, it's acceptable. If you're lucky there is some sensible admin who restricts algorithms.
I'm still worried that this adds too much false sense of security as it is right now.
Encrypting everything everywhere increases the cost of snooping, which is good. It's a long way from perfect, but that shouldn't stop us from making an improvement.
It is putting things in a different envelope for each hop you are doing in the post.
Yes, it is secured in transit, but there is no way that you can verify that everything went securely.
Isn't this the direction newer protocols are taking? AFAIK HTTP 2.0 is always encrypted (even if no SSL cert is installed)
For the mentioned protocols: most IMAP and SMTP servers (which enable TLS) disregard the IETF and allow TLS-on-connect connections on ports 993 and 465 respectively. If the IETF would (again) standardize this the problem would be reasonably solved.
Using ad-hoc certificated is as safe as using STARTTLS: a MITM can just put his own cert (similar to a MITM removing the STARTTLS capability announcement from the client to the server). It's actually worse, since the client will think it's communication is safer than it actually is.
Encryption and identity assurance are two different things - you can easily have the first which makes passive monitoring much more difficult without the trust infrastructure to mitigate MITMing.
For the average user, knowing what we know now, I'd say the first is a much more clear and present danger. The second requires a person to actually take interest in your traffic.
...that defeats the "use TLS when available" argument.