SPN Canonicalization

[Jeffrey Altman]

Christopher Mason wrote:

> Yeah, this is a messy, long-standing, and unresolved issue in kerberos. 
> In theory canonicalization introduces a potential MITM attack which
> essentially undermines mutual authentication; in practice, everyone does
> it.  This is apparently to be addressed in future protocol mods, but
> it's unclear what to do in software that actually needs to work today. 
> Read on for details.  Maybe this could update the kerberos FAQ?

MIT's Kerberos implementation through the 1.4.1 release does indeed use
DNS to resolve the user entered host name and translate it to the
reverse lookup name.  The reason for doing this is that it was assumed
at one point in time that

(a) the DNS infrastructure and the Kerberos infrastructure would be
controlled by the same entity
(b) it is more important to simplify Kerberos so that fewer keys must be
managed on a single machine

The reality is that today (a) is frequently not true and in today's
environment "simple" loses out to "secure".

The MIT Kerberos team has already checked in changes to allow the
DNS-based hostname transformation within the Kerberos libraries to be
disabled.

> I've done some googling on this issue and the clearest statement I can
> yet find comes at the end of section 1.3 in RFC4120, which states:
> 
>> Implementations of Kerberos and protocols based on Kerberos MUST
>> NOT use insecure DNS queries to canonicalize the hostname
>> components of the service principal names (i.e., they MUST NOT
>> use insecure DNS queries to map one name to another to determine
>> the host part of the principal name with which one is to
>> communicate).
> 
> [...]
> 
>> Implementation note: Many current implementations do some degree
>> of canonicalization of the provided service name, often using DNS
>> even though it creates security problems.  However, there is no
>> consistency among implementations as to whether the service name
>> is case folded to lowercase or whether reverse resolution is
>> used.  To maximize interoperability and security, applications
>> SHOULD provide security mechanisms with names that result from
>> folding the user- entered name to lowercase without performing
>> any other modifications or canonicalization.
> 
> 
> Clear as mud?  So it seems like we're left with a choice between
> interoperability and correctness.  Almost all other discussion I find of
> this issue seem to indicate that canonicalization is performed in
> practice.  Firefox and IE definitely do it.  See also question 2.14 in
> the Kerberos FAQ:
> <http://www.cmf.nrl.navy.mil/CCS/people/kenh/kerberos-faq.html#kerbdns>,
> and numerous previous discussions on comp.protocols.kerberos such as
> this thread on load balancing and kerberos:
> <http://mailman.mit.edu/pipermail/kerberos/2004-April/005102.html>.

In the case of a cluster of machines providing services for a given
service, without DNS-based hostname translation the keytab on a
particular machine must contain both the unique machine-specific name as
well as the name given to the cluster.

  host/cluster-machine-1.domain at REALM
  host/cluster.domain at REALM

When DNS-based hostname translation is used, only the unique
machine-specific name must have a key provided.

> Samba seems to work around this problem by creating a skillion different
> keytab entries like:
> 
>   3 HTTP/proteomics.mayo.edu at MFAD.MFROOT.ORG
>   3 HTTP/proteomics at MFAD.MFROOT.ORG
>   3 HTTP/PROTEOMICS at MFAD.MFROOT.ORG
>   3 HTTP/PROTEOMICS.mayo.edu at MFAD.MFROOT.ORG
>   3 HTTP/proteomics.MFAD.MFROOT.ORG at MFAD.MFROOT.ORG
>   3 HTTP/proteomics.mfad.mfroot.org at MFAD.MFROOT.ORG
> 
> and many more besides these (with different salts?).  See
> <http://lists.samba.org/archive/samba-technical/2004-November/038234.html>.
> There's an additional problem with samba, active directory validated
> writes, and disjoint namespaces that I'll save for another time.

Samba is working around a problem that is slightly different from the
multi-homing discussion.  Depending upon how the Microsoft Windows
CIFS/SMB client determines the name for the file server, the case of the
 name will be different and whether or not it is a fully-qualified
domain name will be different.   For a Microsoft Windows server, it does
not care what service name is used.  The server does not contain a list
of service principal names and host keys.   Instead it has a password
and will generate on the fly a key for the service principal name that
is used by the client.  This works well for Windows because Windows
treats names in a case insensitive manner.   All that matters is that
the machine password can be used to generate a key that decrypts the
service ticket.

Samba uses Kerberos libraries that are case sensitive.  Therefore, they
must provide separate keytab entries for each of the expected service
principals that a Windows client joined to a domain might provide.

> It seems to my uneducated eye that really GSSAPI should be such that we
> simply hand it "HTTP" along with the host name entered by the user and
> it figures out the rest.  Alas, it is not so.  What to do is anyone's
> guess.

GSSAPI on the client side does do what it should.  Given a
"service at host" when using a Kerberos mechanism it will obtain a Kerberos
service ticket and establish a connection to the service.

The requirement to make this work is that the names entered by the user
(or those constructed by lookup) must have service principal names in
the KDB and the keytab files on the machines that are hosting the
service must contain entries for all of the service principal names that
might be provided.

Jeffrey Altman



-- 
-----------------
This e-mail account is not read on a regular basis.
Please send private responses to jaltman at mit dot edu