PostgreSQL security releases out

As you can see, PostgreSQL has just released new updated versions, which include security fixes. They also contain other critical bug fixes, so even if you are not directly affected by the security issues, plan an upgrade as soon as possible.

One of the security issues that have been patched deal with NULL prefixes in SSL certificate names, a vulnerability that is basically the same one that have surfaced in a lot of different products this autumn, for example in the Mozilla suite of products. There is not really space enough to properly discuss the implications this has in a PostgreSQL environment in the release notes, so I'll try to elaborate some here - given that I wrote the fix for it.

First of all, a quick explanation of what the problem is. PostgreSQL uses OpenSSL to deal with certificates. Prior to the fixed version, we just asked OpenSSL for the name of the certificate, got back a string, and used this one. Now, if you know C coding, you know that a string is terminated by a NULL character. The bug in PostgreSQL is that we did not check the return value from this function, and make sure it returned the same value as the length of the returned string. This means that somebody could embed a NULL value in the certificate, and we would incorrectly parse and validate only the part that was before the NULL value. For example, if someone managed to get a certificate with the common name set to "postgresql.bank.com\0attacker.com", PostgreSQL would match this certificate against "postgresql.bank.com" (or ".bank.com"), which is not correct. With the fix, the certificate will be rejected* completely.

It is important to know that in order to make use of this vulnerability, the attacker needs to convince a trusted CA to sign such a certificate - which is quite obviously malicious. If the attacker cannot get the CA to hand this out, PostgreSQL will reject the certificate before we even get this far. It is arguably also a bug in the CA handling (technical or procedural) to even hand out such a certificate, and that bug need to be exploited before the one in PostgreSQL can be.

In the vast majority of cases, if not all, where PostgreSQL is deployed and actually using certificate validation, the certificates will be handed out by a trusted local CA. In which case, exploiting this vulnerability becomes much harder. This scenario is significantly different from the original scenario this bug was discovered in, which is the web browser. In the web browser case, the browser already trusts a large number of external CAs by default. PostgreSQL will trust no CAs by default (unless you are doing a debian install, in which case they put some default CAs in there - this is another reason why this is a really bad idea from a security perspective). PostgreSQL also does not prompt the user with a potentially incorrect name field on the certificate asking if this is ok or not - it will just reject the certificate if it doesn't match (correctly or incorrectly), closing another attack venue. So the bug is really only significant if you can't trust your CA - but the whole point of the CA is that it is a trusted entity...

PostgreSQL 8.4 is the first version to properly support certificate name validation, and also the first version to support client certificate authentication, both of which are vulnerable to this bug, neither of which is enabled by default. However, previous versions are also indirectly vulnerable, because they exposed the CN field of the certificate to the application for further validation. So you could have a stored procedure checking the client certificate, or just the libpq application checking the server certificate, even in earlier versions. And given the API structure, there was no way for these outside processes to know if they were being fooled or not. So if you are using an application that makes use of this on previous versions of PostgreSQL, you still need the patch - there is no way to fix the bug from the application.

The summary of this post is that this vulnerability is a lot less serious in PostgreSQL than in many other systems that had the issue. That doesn't mean it's not there, and that it should be (and have been) fixed. But it means that this vulnerability alone is likely not reason enough to rush an upgrade on your production systems - most likely you're not affected by it. On the PostgreSQL security page it is tagged with classification A, which is the highest. This is more an indication that the system we're using for classification really doesn't take these things into consideration - something we will look into for the future.


Comments

Thx for the explanation - and your work.

Regards, Andreas

Posted on Dec 15, 2009 at 06:44 by Andreas Kretschmer.

Add comment

New comments can no longer be posted on this entry.

Conferences

I speak at and organize conferences around Open Source in general and PostgreSQL in particular.

Upcoming

Nordic PGDay 2025
Mar 18, 2025
Copenhagen, Denmark

Past

PGConf.EU 2024
Oct 22-25, 2024
Athens, Greece
PGConf NYC 2024
Sep 30-Oct 2, 2024
New York, USA
PGDay UK 2024
Sep 11, 2024
London, UK
PGConf.DEV 2024
May 28-31, 2024
Vancouver, Canada
PGDay Chicago 2024
Apr 26, 2024
Chicago, USA
More past conferences