Re: [TLS] Summarizing identity change discussion so far

[Kyle Hamilton <[e-mail]>]

On Thu, Dec 17, 2009 at 2:06 PM,  <[e-mail]> wrote:
> Well, the text I proposed just said "different certificate", not
> "memcmp". I know that's a bit vague, but "in accordance with PKIX
> [PKIX]" is not very precise either, and would need a more specific
> reference to where exactly the details are (for example, it can't mean
> comparing the Subject field, since PKIX allows leaving it empty).

PKIX allows (and, actually, encourages) Subject to be empty, but only
in the case where subjectAlternativeName isn't.  There are also cases
where subjectAlternativeName can contain the old Subject, and a new
Subject nym is included in the certificate.

> And the certificate might have extensions (certificate policies,
> subject directory attributes, extended key usage, etc.) where changes
> between API calls could surprise the application.

Then, for those applications, they should be rewritten to terminate
the active session when the policies associated with the certificate
change (more specifically, when a particular eKU that the application
requires in the certificate associated with the connection is no
longer asserted by the peer, when the certificate policies change, or
whenever some condition set by the application developer is no longer
met).  Or, the underlying libraries should enforce this (if asked).
Either way, terminating the connection and invalidating the session is
the easiest way to enforce that all certificate changes come through
the "initial request authenticator/authorizer" code path, which is
where all of those policies are verified.

> Here's another shot at that paragraph (also including suggestions
> from Marsh's email):
>
>   TLS implementations SHOULD offer the applications the option to
>   disable renegotiation completely.
>
>   To make life simpler for applications that do not expect the peer's
>   certificate to change once it's been authenticated, TLS
>   implementations may also wish to offer the applications the option
>   abort the renegotiation if the peer tries to authenticate with a
>   different certificate and/or different server name (in the
>   server_name extension) than was used earlier. However, enabling
>   this option by default for all applications could break existing
>   applications that depend on using renegotiation to change from one
>   certificate to another. (For example, long-lived TLS connections
>   could change to a renewed certificate; or renegotiation could
>   select a different cipher suite that requires using a different
>   certificate.)

I'd agree with that, if it were simply a matter of "certificate" or
"server name".  But, it's not.  This is a matter of identity, and the
identity matching rules of PKIX -- since PKIX is what defines the
content and schema of the X.509 certificates presented by both pairs
-- are what must control, not simply any arbitrary decision by the TLS
WG.  Nor can even any decision made by the TLS WG chair control.

(Certificate type 0 is X.509/PKIX.  Certificate type 1 is OpenPGP --
oh, but wait, we don't have anyone here to advocate the matching rules
for its particular IANA assignment... Each type, 0 and 1, has
different matching rules.  The rule for verifying the token that is
presented MUST be defined, in TLS, as the rule defined by the document
that defines the format of the token, or else the TLS WG is once again
putting a policy decision into a technical specification, restricting
the kinds of authentication which can be used.)

-Kyle H