[kitten] Review of draft-ietf-kitten-channel-bound-flag-04

[Sam Hartman <[e-mail]>]

I've reviewed this document at the request of one of the authors, and
here are my comments:

This draft is based on the premis that the GSS-API channel binding
behavior is flawed and that  there's a security problem associated with
not knowing if channel bindings are validated.

I've heard this argument over the years, but never been convinced of it,
and unfortunately, this draft is filled with assertions of that fact
without any rationale.

At one level it doesn't matter much.  I absolutely agree that a facility
for knowing if channel bindings are validated would be incredibly useful
and would simplify application design.

However, by jumping to the conclusion that we're fixing a flaw, I think
we approach interoperability and operational aspects with inadequate
thought.  I think we discard some of the options too early.

My past experience is that when we take short cuts like that, we find
ourselves frustrated later when we discover things we overlooked.

As I write this up, I'm realizing that the draft authors seem to be
making an assumption that  there are applications in the wild that pass
in channel bindings to the initiator but not the acceptor or vise
versa.  The only such application I'm aware of is gss ftp, which passed
IP address channel bindings into the acceptor and optionally initiator.
It would help me evaluate whether there is a security problem to
understand these applications better.
I note that SASL is not such an application.

Regardless of whether this is a security problem, I am very supportive
of a mechanism along these lines.  I think that it would simplify
application design over the approaches necessary for the gs2 family of
mechanisms.

I think the current draft has several shortcomings:

* I think that it inaccurately describes what RFC 2743 says and so it
does not call out backward incompatible changes.

* I think that it inaccurately describes what current implementations
  do.

* I think that more consideration needs to be given to situations where
one side supports the channel bound flag and the other does not and what
if any signaling changes are required.

* I think it does not consider behavior when mutual authentication is
not requested.

* I'm concerned that in thinking about channel binding we have not given
adequate thought to how channel binding interacts with proxies, reverse
proxies and the like.  I'm concerned that this draft makes that
situation worse, and so I'd recommend exploring that at least enough to
understand whether I'm right.

* Please get review from multiple implementors before standardizing the
create_context stuff.  In particular, do you want to take inputs like a
credential and/or a flag indicating the context is for initiation or
acceptor.  If yo,u've gotten review including from folks like Luke and
they are happy, then this is great.  It seems good for API users at
least.

Here are my specific comments:




>Note that it would be useless to allow security context
>   establishment to succeed when the initiator does not provide channel
>   bindings but the acceptor does, at least as long as there's no
>   outward indication of whether channel binding was used!

An exclamation in an IETF document is a good sign that you're letting
your emotions get ahead of your technical content and perhaps you should
take a step back.
I think the above quote is a good example.
You don't get MITM protection if the initiator provides channel binding
and the acceptor does, but it can be useful none-the-less.
Assuming that an attacker cannot convert a context token that has
channel bindings into one that does not, and initiators know acceptors
are providing channel bindings, then initiators know that if they
provide channel bindings then the channel bindings are validated.
That can provide you defense in depth, for example, binding channels to
a TLS channel binding in applications where you also validate the PKI
and have good trust anchors.
It cannot be sufficient alone to provide MITM defense.
However, additional layers of defense (provided they do not trigger
false negatives) are valuable in security.
Please rephrase to focus on the characteristics of the system rather
than judgments of those characteristics.

>   o  Whenever both the initiator and acceptor have provided
>      input_channel_bindings to GSS_Init/Accept_sec_context() and the
>      channel bindings do not match, then the mechanism MUST fail to
>      establish a security context token.  (This is a restatement of an
>      existing requirement in the base specification.)

Actually, as best I can tell, RFC 2743section 1.1.6 disagrees with
this.  It looks in RFC 2743 like channel bindings are optional for a
mechanism to use.
I think it's reasonable that if a mechanism supports channel bindings,
then if both the initiator and acceptor provide channel bindings they
must validate for context establishment to succeed.

>   o  Whenever the acceptor application has a) provided channel bindings
>      to GSS_Accept_sec_context(), and b) not indicated support for the
>      ret_channel_bound_flag flag, then the mechanism MUST fail to
>      establish a security context if the initiator did not provide
>      channel bindings data.  This requirement is critical for security
>      purposes, to make applications predating this document secure, and
>      this requirement reflects actual implementations as deployed.

Actually, this behavior is inconsistent with RFC 7055 section 5.6.2 and
with the Moonshot implementation of the ABFAB GSS-API mechanism.
In that mechanism, if the initiator provides null channel bindings, the
acceptor channel bindings are always ignored.
I'm not seeing the security issue here.  If you are using channel
bindings as a MITM defense rather than as an additional layer of
security, you need to do something like SASL does and have a plus and
minus varient of your protocol and always pass in channel bindings.

There's also what I think is a real operational issue here.  You may not
be able to know what your channel bindings are on an acceptor.
Depending on whether you are behind a TLS accelerator, reverse proxy or
the like, it may affect your ability to know your channel bindings.

I
>   o  Whenever the application has a) provided channel bindings to
>      GSS_Init_sec_context() or GSS_Accept_sec_context(), and b)
>      indicated support for the ret_channel_bound_flag flag, then the
>      mechanism SHOULD NOT fail to establish a security context just
>      because the peer did not provide channel bindings data.  The
>      mechanism MUST output the ret_channel_bound_flag if both peers
>      provided the same input_channel_bindings to GSS_Init_sec_context()
>      and GSS_Accept_sec_context().  The mechanism MUST NOT output the
>      ret_channel_bound_flag if either (or both) peer did not provide
>      input_channel_bindings to GSS_Init/Accept_sec_context().  This
>      requirement restores the original base GSS-API specified behavior,
>      with the addition of the ret_channel_bound_flag flag.

I think we need to consider the transition period here.
What happens when you have a new initiator talking to an old acceptor or
vice versa.Are you planning on introducing new over-the-wire signaling
so that an acceptor  can say something like "I'm setting up your
context, but your channel binding didn't verify?"
For example in RFC 7055, the channel binding data is sent as a checksum
from initiator to acceptor.  The acceptor either sends back a final
token or sends back an error depending on whether it accepts the
bindings.
How, without protocol changes would we meet this requirement.
Similarly, I think Kerberos works the same way in the mutual
authentication case.  In the non mutual authentication case, I don't
think the initiator even learns whether the acceptor validated the
channel bindings.