Re: [TLS] Revision 10: possible attack if client authentication is allowed during PSK

[Karthikeyan Bhargavan <[e-mail]>]

Nice analysis! I think that the composition of different mechanisms in the protocol is likely to
be where many subtle issues lie, and analyses like this one support that concern.

Several other composition examples were already discussed on this list (for previous drafts).
For example, when composing 0-RTT with 1-RTT in the most natural way, many attacks appeared:
   (1) 0-RTT application data  may be replayed (See note (2) in 6.2.2.)
   (2) an unknown key share attack on 0-RTT encryption keys (leading to the inclusion of ServerCertificate in 0-RTT handshake hash in 7.2.1)
   (3) a key compromise impersonation attack on 0-RTT (See note (3) in 6.2.2)

The addition of resumption, PSK, and client-auth provide plenty of further composition possibilies
that are (in my surely biased opinion) best handled by automated formal analysis. 
So, I’m looking forward to the full Tamarin results!

Best,
Karthik

> On 31 Oct 2015, at 12:19, Sam Scott <[e-mail]> wrote:
> 
> Dear all,
> 
> We [1] are in the process of performing an automated symbolic analysis
> of the TLS 1.3 specification draft (revision 10) using the Tamarin
> prover [2], which is a tool for automated security protocol analysis.
> 
> While revision 10 does not yet appear to permit certificate-based client
> authentication in PSK (and in particular resumption using PSK), we modelled
> what we believe is the intended functionality. By enabling client
> authentication either in the initial handshake, or with a post- handshake
> signature over the handshake hash, our Tamarin analysis finds an attack. The
> result is a complete breakage of client authentication, as the attacker can
> impersonate a client when communicating with a server:
> 
> Suppose a client Alice performs an initial handshake with Charlie. Charlie,
> masquerading as Alice, subsequently performs a handshake with Bob. Following a
> PSK resumption, Bob requests authentication from Charlie (impersonating Alice).
> Charlie then requests authentication from Alice, and the returned signature
> will also be a valid signature for the session with Bob.
> 
>         Initial h/s                                      Initial h/s
>      |<-------------->|                               |<-------------->|
>      |  exchange PSK  |                               |  exchange PSK  |
>      |                |                               |                |
>      |Start PSK resume|                               |Start PSK resume|
>      |--------------->|                               |--------------->|
>      |client_random nc|                               |client_random nc|
>      |                |                               |                |
>      |  Accept resume |                               |  Accept resume |
> Alice|<---------------|(as Charlie) Charlie (as Alice)|<---------------|Bob
>      |server_random ns|                               |server_random ns|
>      |                |                               |                |
>      |                |                               |                |
>      |Client auth req |                               |Client auth req |
>      |<---------------|                               |<---------------|
>      |                |                               |                |
>      |  Client auth   |                               |  Client auth   |
>      |--------------->|                               |--------------->|
>        sign nc,ns,...                                  relay signature
> 
> 
> 
> This attack is possible because the client signature is over the handshake
> hash, which only covers the nonces and other easily duplicated information,
> and in particular does not contain the server certificate (because none is
> presented in PSK mode [3]) or the server Finished.
> 
> While the modifications proposed in PR#316 [4] explicitly allow client
> authentication in these contexts, the PR also redefines the client signature
> based on a new "Handshake Context" value which includes the server Finished
> message. Intuitively, this new definition appears to address the attack
> because the attacker cannot transplant the Finished message between
> connections. We are currently working towards a Tamarin proof that PR#316
> indeed prevents our attack.
> 
> Therefore we would like to support the inclusion of Finished as
> part of the handshake context, in order to address this problem.
> 
> Many thanks,
> 
> Sam Scott
> 
> [1] Cas Cremers, Marko Horvat - University of Oxford;
>     Thyla van der Merwe, Sam Scott - Royal Holloway, University of London.
> [2] http://www.infsec.ethz.ch/research/software/tamarin.html <http://www.infsec.ethz.ch/research/software/tamarin.html>
> [3] Except in 0-RTT where the server's Certificate is explicitly
>     included.
> [4] https://github.com/tlswg/tls13-spec/pull/316/ <https://github.com/tlswg/tls13-spec/pull/316/>_______________________________________________
> TLS mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/tls