Re: [TLS] Proposal for detecting fraudulent certificates

[Martin Rex <[e-mail]>]

Florian Weimer wrote:
> 
> I have submitted draft-weimer-tls-previous-certificate-00, which intents
> to facilitate detection of fraudulent certificates used in the wild:
> 
> <http://tools.ietf.org/html/draft-weimer-tls-previous-certificate-00>
> 
> The basic idea is to use leaks from mobile clients moving between
> networks with and without a clear path to the server.  The previous
> server certificate chain is included in the client hello, so the server
> receives it when the client transitions to a network with a clear path.
> 
> This draft prompted my previous question about extension size limits.
> Unfortunately, that issue makes this very simple idea somewhat
> complicated, but I tried to add a fairly straightforward workaround.

Having the client send _only_ the servers certificate and not the
entire certificate path should IMO be sufficient for the purpose you
outline.

I don't know why you think the 64 KByte TLS extension size limit would
be a problem.

A non-marginal fraction of the installed base of TLS will puke when
encountering a TLS handshake message that is split/fragmented across
TLS record boundaries (which have a limit of 16 KByte).  The currently
most likely situation where this occurs (and the only one that
I've seen in the wild so far) are the CertificateRequest handshake
message contains a large number (60+) of certification_authorities.
Usually, the server tries to stuff the server response
(ServerHello,Certificate, CertificateRequest,ServerHelloDone)
into one single 16 KByte SSL Record, so even fitting the Server's
certificate chain alone should not normally be a problem.


But I don't see how this extension could realistically work in a
sane fashion.  The TLS implementation in TLS servers is not normally
keeping a history of previous TLS Server certificates which it ever
used (and might not have a persistence of its own for such a purpose),
usually not even across process restarts, and having TLS backends of
a Web-Server-Farm or hot-backups keep such a history in a consistent
fashion is even more unlikely.

And end users would probably prefer their TLS clients to tell them right
away when a server credential changes unexpectedly, rather than
a TLS server telling them, after they come home from a several weeks
trip through some contry YYY, that their last several week of connect(s)
had all been MITMd...


-Martin