Websphere cookie/sessionid predictable

From: Marc Heuse (marc@suse.de)
Date: 09/19/01


Subject: Websphere cookie/sessionid predictable
To: bugtraq@securityfocus.com
Date: Wed, 19 Sep 2001 19:43:39 +0200 (CEST)
Message-Id: <20010919174339.B8DF01FBB3@Appserv.suse.de>
From: marc@suse.de (Marc Heuse)

Hi folks,

about three weeks ago, I discovered a hole in IBM's websphere 4.0 session ID
generation. Over a week ago, IBM made a fix for this available, so here is
the information about the vulnerability:

(everybody who don't want to read about this vulnerability and just want to
know the patch info: install the eFix PQ47663V302)

INTRO
websphere can generate sessionids which are put into cookies for users, to
be able to supply user tracking, e.g. user authenticates with userid and
password, and access to data is checked by checking if the sessionid is
authenticated or not.

THE BUG
during a security assessment for a bank, I collected several sessionids and
they did not look that random to me ...

SessionID TIME
TWGYLZIAAACVDQ3UUSZQV2I 10:27:12
TWGY0WYAAACVFQ3UUSZQV2I 10:27:13
TWGZNZAAAACVHQ3UUSZQV2I 10:27:14
TWG0BUYAAACVJQ3UUSZQV2I 10:27:15
TWG0VIAAAACVLQ3UUSZQV2I 10:27:16
TWG1ICIAAACVNQ3UUSZQV2I 10:27:17
TWG111YAAACVPQ3UUSZQV2I 10:27:18
   xxxx y

You can see that for seven requests, only 5 characters changed, and:
* the characters A-Z and 0-9 are used, hence 36 combinations possible per char
* the sessionid is based on two counters which are counted up, the rest of
  the string seems to be fixed.
* the first counter (xxxx) seems to count milliseconds (TWGxxxx), but
  counts a bit too slow (see seconds 15 and 16, where both 1st rows of the
  counters start with a 0). well, to cut a long story short, it is really
  a counter which increases 65536 times per second and is then encoded to
  the A-Z0-9 format.
* If you collect many sessionids (I collected 1000, it's obvious then),
  you'll see that the the least signifcant char of the first counter are 95%
  of the time showing a Y, I, A or Q. The reason for that is (my guess) that
  the clock of the machine only can increase 7.500-10.000 times instead of
  65536 because it's not a realtime clock and the server type is not a cray :-)
* The second counter (y) is increasing by two every second.
The counters are in fact longer than 4, but this is the visible changes in
the example above.

Then the guess was that the fixed strings may be based on the IP of the
client. So I checked with different IP addresses, but no difference in the
fixed strings of the sessionID.

THE RISK
If someone knows the time of the connect to the server (even with SSL
encrypted) an attacker can issue requests with changing sessionids until
it's the correct one. If an attacker just wants to have any user data, he
can constantly try some guessing.
As the first counter only has 7.500-10.000 values per second, and the
seconds counters just increases approx. once per second (or perhaps per
request), the sessionid can have 7.750 to 10.500 different values.
If a user is normaly connected for 15 minutes after authentication to an
eCommerce system (and does not forget to logout, otherwise the time is
extended by the session timeout). As an attacker is likely to succeed after
50% of the keyspace, he needs 3.875 to 5.250 attempts, so 4 to 5 requests
per second are enough.
Two customers were using the sessionids for the security of their eCommerce
system ... we are not talking about some weird feature nobody uses.
Short: it is an easy and likely attack.

THE FIX
install eFix PQ47663V302 and feel better

THANKS
to the IBM websphere team, which fixed the bug pretty fast for the customer.

Greets,
        Marc

--
   E@mail: marc@suse.de  Function: Security Research and Advisory
  PGP: "lynx -source http://www.suse.de/~marc/marc.pgp | pgp -fka"
 Key fingerprint = B5 07 B6 4E 9C EF 27 EE  16 D9 70 D4 87 B5 63 6C
Private: http://www.suse.de/~marc  SuSE: http://www.suse.de/security