OpenSSL Security Advisory [19 February 2003]

Timing-based attacks on SSL/TLS with CBC encryption
===================================================

CONTENTS

 - Vulnerability
 - Source code patch [*]
 - Acknowledgement
 - References

[*] OpenSSL 0.9.6i and OpenSSL 0.9.7a do not require this patch.

The source code of OpenSSL 0.9.6i and OpenSSL 0.9.7a is available as
files openssl-0.9.6i.tar.gz and openssl-0.9.7a.tar.gz from

     ftp://ftp.openssl.org/source;type=d

(If you were previously using an OpenSSL 0.9.6* "engine" release and
cannot upgrade to OpenSSL 0.9.7a, obtain file
openssl-engine-0.9.6i.tar.gz instead.  With OpenSSL 0.9.7, the
"engine" framework has become part of the standard distribution.)

If you are using a pre-compiled OpenSSL package, please look for update
information from the respective software distributor.  The OpenSSL
group itself does not distribute OpenSSL binaries.


Vulnerability
-------------

In an upcoming paper, Brice Canvel (EPFL), Alain Hiltgen (UBS), Serge
Vaudenay (EPFL), and Martin Vuagnoux (EPFL, Ilion) describe and
demonstrate a timing-based attack on CBC ciphersuites in SSL and TLS.

The attack assumes that multiple SSL or TLS connections involve a
common fixed plaintext block, such as a password.  An active attacker
can substitute specifically made-up ciphertext blocks for blocks sent
by legitimate SSL/TLS parties and measure the time until a response
arrives: SSL/TLS includes data authentication to ensure that such
modified ciphertext blocks will be rejected by the peer (and the
connection aborted), but the attacker may be able to use timing
observations to distinguish between two different error cases, namely
block cipher padding errors and MAC verification errors.  This is
sufficient for an adaptive attack that finally can obtain the complete
plaintext block.

OpenSSL version since 0.9.6c supposedly treat block cipher padding
errors like MAC verification errors during record decryption
(see http://www.openssl.org/~bodo/tls-cbc.txt), but MAC verification
was still skipped after detection of a padding error, which allowed
the timing attack.  (Note that it is likely that other SSL/TLS
implementations will have similar problems.)

OpenSSL 0.9.6i and 0.9.7a perform a MAC computation even if incorrrect
block cipher padding has been found to minimize information leaked via
timing.  For earlier versions starting with 0.9.6e, the enclosed
security patch can be used.


Source code patch
-----------------

If upgrading to OpenSSL 0.9.7a (the recommended version) or to OpenSSL
0.9.6i is not immediately possible, the following patch should be
applied to the OpenSSL source code tree.  The patch is compatible
with OpenSSL 0.9.6e and later.

>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>
--- ../openssl-0.9.6h/CHANGES  Thu Dec  5 22:40:48 2002
+++ ./CHANGES  Tue Feb 19 00:00:00 2003
@@ -1,3 +1,19 @@
+ Change from security patch  [19 Feb 2003]
+
+ (Please consider installing OpenSSL 0.9.7a or OpenSSL 0.9.6i.
+ These versions already include this change; do not try to apply
+ the patch to them!)
+
+  *) In ssl3_get_record (ssl/s3_pkt.c), minimize information leaked
+     via timing by performing a MAC computation even if incorrrect
+     block cipher padding has been found.  This is a countermeasure
+     against active attacks where the attacker has to distinguish
+     between bad padding and a MAC verification error. (CAN-2003-0078)
+
+     [Bodo Moeller; problem pointed out by Brice Canvel (EPFL),
+     Alain Hiltgen (UBS), Serge Vaudenay (EPFL), and
+     Martin Vuagnoux (EPFL, Ilion)]
+
 
  OpenSSL CHANGES
  _______________
--- ../openssl-0.9.6h/ssl/s3_pkt.c  Mon May  6 12:42:56 2002
+++ ./ssl/s3_pkt.c  Wed Feb 18 00:00:00 2003
@@ -238,6 +238,8 @@
   unsigned int mac_size;
   int clear=0;
   size_t extra;
+  int decryption_failed_or_bad_record_mac = 0;
+  unsigned char *mac = NULL;
 
   rr= &(s->s3->rrec);
   sess=s->session;
@@ -353,8 +355,11 @@
       /* SSLerr() and ssl3_send_alert() have been called */
       goto err;
 
-    /* otherwise enc_err == -1 */
-    goto decryption_failed_or_bad_record_mac;
+    /* Otherwise enc_err == -1, which indicates bad padding
+     * (rec->length has not been changed in this case).
+     * To minimize information leaked via timing, we will perform
+     * the MAC computation anyway. */
+    decryption_failed_or_bad_record_mac = 1;
     }
 
 #ifdef TLS_DEBUG
@@ -380,28 +385,46 @@
       SSLerr(SSL_F_SSL3_GET_RECORD,SSL_R_PRE_MAC_LENGTH_TOO_LONG);
       goto f_err;
 #else
-      goto decryption_failed_or_bad_record_mac;
+      decryption_failed_or_bad_record_mac = 1;
 #endif      
       }
     /* check the MAC for rr->input (it's in mac_size bytes at the tail) */
-    if (rr->length < mac_size)
+    if (rr->length >= mac_size)
       {
+      rr->length -= mac_size;
+      mac = &rr->data[rr->length];
+      }
+    else
+      {
+      /* record (minus padding) is too short to contain a MAC */
 #if 0 /* OK only for stream ciphers */
       al=SSL_AD_DECODE_ERROR;
       SSLerr(SSL_F_SSL3_GET_RECORD,SSL_R_LENGTH_TOO_SHORT);
       goto f_err;
 #else
-      goto decryption_failed_or_bad_record_mac;
+      decryption_failed_or_bad_record_mac = 1;
+      rr->length = 0;
 #endif
       }
-    rr->length-=mac_size;
     i=s->method->ssl3_enc->mac(s,md,0);
-    if (memcmp(md,&(rr->data[rr->length]),mac_size) != 0)
+    if (mac == NULL || memcmp(md, mac, mac_size) != 0)
       {
-      goto decryption_failed_or_bad_record_mac;
+      decryption_failed_or_bad_record_mac = 1;
       }
     }
 
+  if (decryption_failed_or_bad_record_mac)
+    {
+    /* A separate 'decryption_failed' alert was introduced with TLS 1.0,
+     * SSL 3.0 only has 'bad_record_mac'.  But unless a decryption
+     * failure is directly visible from the ciphertext anyway,
+     * we should not reveal which kind of error occured -- this
+     * might become visible to an attacker (e.g. via a logfile) */
+    al=SSL_AD_BAD_RECORD_MAC;
+    SSLerr(SSL_F_SSL3_GET_RECORD,SSL_R_DECRYPTION_FAILED_OR_BAD_RECORD_MAC);
+    goto f_err;
+    }
+
   /* r->length is now just compressed */
   if (s->expand != NULL)
     {
@@ -443,19 +466,12 @@
 
   return(1);
 
-decryption_failed_or_bad_record_mac:
-  /* Separate 'decryption_failed' alert was introduced with TLS 1.0,
-   * SSL 3.0 only has 'bad_record_mac'.  But unless a decryption
-   * failure is directly visible from the ciphertext anyway,
-   * we should not reveal which kind of error occured -- this
-   * might become visible to an attacker (e.g. via logfile) */
-  al=SSL_AD_BAD_RECORD_MAC;
-  SSLerr(SSL_F_SSL3_GET_RECORD,SSL_R_DECRYPTION_FAILED_OR_BAD_RECORD_MAC);
 f_err:
   ssl3_send_alert(s,SSL3_AL_FATAL,al);
 err:
   return(ret);
   }
+const char *CAN_2003_0078_patch_ID="CAN-2003-0078 patch 2003-02-19";
 
 static int do_uncompress(SSL *ssl)
   {
<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<


Acknowledgement
---------------

We thank Brice Canvel, Alain Hiltgen, Serge Vaudenay, and Martin
Vuagnoux for informing us about this vulnerability.


References
----------

LASEC Memo "Password Interception in a SSL/TLS Channel" by Brice Canvel,
published 2003-02-20:
http://lasecwww.epfl.ch/memo_ssl.shtml

The Common Vulnerabilities and Exposures project (cve.mitre.org) has
assigned the name CAN-2003-0078 to this issue:
http://cve.mitre.org/cgi-bin/cvename.cgi?name=CAN-2003-0078

URL for this Security Advisory:
http://www.openssl.org/news/secadv_20030219.txt