North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

Re: OT - 3 Free Gmail invites

  • From: Eric Rescorla
  • Date: Thu Aug 19 12:17:07 2004

Bill Woodcock <[email protected]> writes:

>       On Thu, 19 Aug 2004, Steven S. wrote:
>     > I have 5 invites that I'm willing to part with...
> 
> Uh, could we _please_ get back to something with operational content, or
> nothing at all?
> 
> Anyone have anything concrete on the SHA-0 / MD5 compromise, for instance?
> Any operational impact there, that we need to worry about in the near
> term?

Here's the overview I sent to IAB/IESG:

As you may or may not have heard, this year's CRYPTO conference
has been very interesting:

* Joux has found a single collision in SHA-0--an algorithm that nobody
  uses but that is very similar to SHA-1. However, SHA-0 was changed to
  fix a flaw (later found by Joux), thus becoming SHA-1 so we can hope
  that this attack can't be extended to SHA-1. The attack was fairly
  expensive, requiring about 2^51 operations the brute force attack
  would take about 2^80).

* Biham and Chen can find collisions in a reduced round version of SHA-1
  (40 rounds). The full SHA-1 is 80 rounds. It's hard to know whether
  this can be extended to full SHA-1 or not. NSA (who designed SHA-1)
  seems to be generally pretty good at tuning their algorithms so that
  they're just complicated enough to be secure.

* Weng, Fang, Lai, and Yu have what appears to be a general method for
  finding collisions in MD4, MD5, HAVAL-128, and RIPEMD. They
  haven't published any details.

What does this mean for us? I'll be writing up full details hopefully
soon, but here's a short overview...

WHAT'S BEEN SHOWN?
An attacker can generate two messages M and M' such that Hash(M) = Hash(M').
Note that he cannot (currently) generate a message M such that Hash(M)
is a given hash value, nor can he generate a message M' such that it hashes
the same as a fixed message M. Currently this is possible for MD5 
but we have to consider the possibility that it will be eventually
possible for SHA-1.


USES OF HASH FUNCTIONS
We use hash algorithms in a bunch of different contexts. At minimum:

1. Digital signatures (you sign the hash of a message).
   (a) On messages (e.g. S/MIME). 
   (b) On certificates.
   (c) In authentication primitives (e.g., SSH)
2. As MAC functions (e.g. HMAC)
3. As authentication functions (e.g. CRAM-MD5)
4. As key generation functions (e.g. SSL or IPsec PRF)

THE POTENTIAL ATTACKS
The only situation in which the current attacks definitely apply is 
(1). The general problem is illustrated by the following scenario.
Alice and Bob are negotiating a contract. Alice generates two
messages:

M  = "Alice will pay Bob $500/hr"
M' = "Alice will pay Bob $50/hr" [0]

Where H(M) = H(M').

She gets Bob to sign M (and maybe signs it herself). Then when it
comes time to pay Bob, she whips out M' and says "I only owe
$50/hr", which Bob has also signed (remember that you sign the
hash of the message). 

So, this attack threatens non-repudiation or any kind of third
party verifiability. Another, slightly more esoteric, case is
certificates. Remember that a certificate is a signed message
from the CA containing the identity of the user. So, Alice
generates two certificate requests:

R  = "Alice.com, Key=X"
R' = "Bob.com, Key=Y"

Such that H(R) = H(R') (I'm simplifying here). 

When the CA signs R, it's also signing R', so Alice can present
her new "Bob" certificate and pose as Bob. It's not clear that
this attack can work in practice because Alice doesn't control
the entire cert: the CA specifies the serial number. However,
it's getting risky to sign certs with MD5.


WHAT'S SAFE?
First, anything that's already been signed is definitely safe.  If you
stop using MD5 today, nothing you signed already puts you at risk.

There is probably no risk to two party SSH/SSL-style authentication
handshakes.

It's believed that HMAC is secure against this attack (according to Hugo
Krawczyk, the designer) so the modern MAC functions should all be
secure.

I worry a bit about CRAM-MD5 and HTTP Digest. They're not as well
designed as HMAC and you might potentially be able to compromise them to
mount some kind of active cut-and-paste attack, though I don't have one
in my pocket.

The key generation PRFs should be safe.

-Ekr


[0] In practice, the messages might not be this similar, but there
turn out to be lots of opportunities to make subtle changes in any
text message.