Wednesday, September 18, 2013

Certificate Transparency and PreCertificates, how will that work?

The Certificate Transparency initiative (RFC6962) is an admirable suggestion to improve security of TLS web session for certificates issued by public CAs. It has cool technology with Merkle trees, is admirable short and could have been straight forward was it not for something called PreCertificates. PreCertificates are hard for me to understand, I don't like them. I hope it is because I don't understand them...if so please let me know.

Writing this post is a way to sort things out for myself and I'd be happy to edit this post if explained why I "just don't get it". Of course I am posting this to the CT forum as well...

In the sake of transparency, I'm writing with the view point of an implementer of open source CA software (if you didn't figure that one out from the blog name:-)).

Update 1: I got lots of comments already over at the Certificate Transparency Forum, really good.

Update 2: I created an issue in the Certificate Transparency issue tracker.

Update 3: Of course my views on CT changes as the discussion continues, the post below was my original starting point. Follow the discussion in Update 1 for updates.

Update 4: EJBCA will support Certificate Transparency in EJBCA Enterprise eventually, watch out for news.

On to PreCertificates...

PreCertificates are defined in section "3.1. Log Entries" as (text trimed by me) "The Precertificate is constructed from the certificate to be issued by adding a special critical poison extension to the end-entity TBSCertificate". Then it describes how it can be produced and it is mentioned throughout the spec in many places.
A PreCertificate is a essentially a certificate signed with one of two options:

1. PreCertificates signed by the real CA.
This sounds very dangerous as will break the fundamental X.509 rule of unique issuerDN/serialNumber pairs. The consequences of having two "certificates" with the same issuerDN/serialNumber in the wild can not possibly be estimated, making this practice quite dangerous imho.

2. PreCertificates signed by a separate PreCertificate signing CA, which is a SubCA to the real signing CA. This is a less scary, since it is normal practice that different CAs can issue certificate with the same subjectDN/serialNumber, just not the same issuerDN.

The actual implementation of issuing PreCertificates makes it quite impractical. I would believe that most CA implementations creates the TBSCertificate as part of the actual certificate issuance. The CA will not create the TBSCertificate to have is lying around for a couple of days before using it to issue the real certificate.
Thus, if the CA is to create a PreCertificate to send to the CT log, it might as well issue the real certificate and send it to the log. The time difference should be in the milliseconds for most CAs.
If the CA wants to wait before distributing the real certificate, to make sure it's in the logs before put into production, it can surely do so as well.

The PreCertificate imho suffers from several complicating factors for implementers, both on the CA and the CT log side. The TBSCertificate must have a poison extension inserted, and removed, effectively re-encoding the ASN.1 TBSCertificate several times, all these are points of failure.

The reason for PreCertificates are not clearly explained. Why would you want to use PreCertificates?

Fine combing through the spec gives me some ideas on why, for example to be able to embed the Certificate extension from PreCertificate CT logs in the final certificate (section 3.3). But the the TBSCertificate of the PreCertificate is then no longer the real TBSCertificate? In that case, why is the PreCertificate the TBSCertificate at all, and not just a new data structure with the data the CT log wants?

The PreCertificate complicates the CT spec by orders of magnitude, which is not a good thing. There are so many ifs and buts about PreCertificate the RFC is not even itself consitent about what it is.

Ok, I know the PreCertificate is is optional, but the best standards, who gets fast, wide and robust deployment, are the simpler ones (KISS). Skipping PreCertificates from the CT spec makes it so much simpler.

My suggestion:
- Skip PreCertificates altogether

I see though why people will not accept that just because I say in that case

- Explain the purpose behind PreCertificates well
- Describe what the actual information fro PreCertificate are used
- Be consistent throughout in the RFC

Feel free to contact me at tomas a t primekey dot se.

No comments: