Re: BUG #17760: SCRAM authentication fails with "modern" (rsassaPss signature) server certificate

On Fri, Feb 10, 2023 at 02:24:12AM +0200, Heikki Linnakangas wrote:
> Note that the discussion in that thread is in a slightly different context.
> They are not talking about channel binding, but something else
> (https://github.com/openssl/openssl/issues/15477#issuecomment-852928702).
> The problem is similar: what hash algorithm to use if it's not well defined.
> But there's no guarantee that the rules they come up with will be the same
> as the rules for tls-server-end-point. So I'm not very comfortable relying
> on X509_digest_sig() to do the right thing for us in the future, even if it
> does today.

Okay.  Yes, it does not seem like a good idea in the long run to avoid
that.

>>> How will this evolve when SCRAM-SHA-512 is implemented? Do you plan to
>>> upgrade the undef-hash in that case to SHA-512?
> 
> Yes, that's what I had in mind. Or for SCRAM-SHA-512, we could go further
> and define it as "always use SHA-512", because we wouldn't need to worry
> about backwards compatibility anymore.

Hmm.  There is a risk of breakage here with older versions of libpq
because we'd still need to support both 256 and 512, no?  Or do you
mean to drop SCRAM-SHA-256 and do a drop-in replacement with
SCRAM-SHA-512?  Doing that in the server may be fine, not in libpq.
As long as you control that with the channel binding name, there
should be enough control I guess.

> Let's define a new channel binding type: tls-server-end-point-sha-256. with
> "tls-server-end-point-sha-256", the SHA-256 hash of the certificate is used,
> regardless of what algorithm is in the certificate's signature.
> 
> 1. When the client sees a "normal" server certificate that uses SHA-256,
> SHA-512, or something else that it recognizes and handles fine today, it
> does the same thing as it does today. It sends "tls-server-end-point" as the
> the cb-name, and calculates the hash using the certificate's hash algorithm.
> 
> 2. When the client sees a certificate with RSA-PSS, ED25519, ED448, or any
> other signature method that it doesn't recognize, it selects the
> tls-server-end-point-sha-256 channel binding type. It sends
> "tls-server-end-point-sha-256" as the cb-name, and calculates the SHA-256
> hash of the certificate.
> 
> 3. The server follows the current rules, or uses SHA-256, depending on
> whether the client sent "tls-server-end-point" or
> "tls-server-end-point-sha-256" in the cb-name.
> 
> These rules ensure that with certificates that work today, nothing changes.
> Rule 1 ensures that you can connect with a new client to an old server, and
> it still works. And rule 3. ensures that you can still connect to a new
> server with an old client.
> 
> At the same time, we are using a new channel binding type to make it clear
> when we are deviating from the standard rules of "tls-server-end-point".

So, using a new channel binding name with tls-server-end-point-sha-256
is actually the way to violate the RFC, while tls-server-end-point
still tries to stick with it?  I am not sure if there's much benefit
in a second channel binding, enforcing SHA-256 for undefined
signatures could just be better, and simpler. :)

There be a risk of a downgrade attack here, where the client could
feign the server to do a SHA-256 computation where SHA-512 or higher
should be used.  Or, if the server certificate has a hash function but
the client feigns in not knowing that, do you mean to go through a
mock authentication in this case and fail?

> If you have an RSA-PSS certificate, you need to upgrade the client and the
> server, or you get an error when you try to connect. (Unless you set
> channel_binding=disable). But after upgrading, it will work.

Indeed.
--
Michael