[ SOURCE: http://www.secureroot.com/security/advisories/9737032598.html ] -----------------.---------------------------------------------. /| | . | / | : : : : : : | / | :: ------ :: : :: | :: - |----- | | :: : :: . : | | :: : | | | : . |------| | : | | | ------^ : | / | . | ;----------"---------------^------ / ------'--------------------- | / / / /----' / / |'----------'---------------'------' --------'---------------------' www.f8labs.com [ INTRODUCTION ] Advisory .........: Nasty tricks with ManTrap Release Date .....: 11-01-00 Application ......: ManTrap by Recourse Technologies Vendor Web Site ..: www.recourse.com Version ..........: All versions prior to and including the newest version 1.6.1 Vendor Status ....: Contacted By ...............: Fate Research Labs WWW ..............: www.f8labs.com [ OVERVIEW ] ManTrap's intent is to set up a honeypot as a 'cage' which can be filled with fake information running on a system where the intruder can be lured and studied, and which is also supposed to prevent erasing of logs etc. A great idea - but the implementation is poor. Instead of running as a VMware like virtual machine or such, it just runs as a chroot like environment with some kernel patches, and thus it can easily be identified and subverted. In case an intruder breaks in and finds out that it's a cage, he/she will most likely just trash it and move on instead of fulfilling the purpose of staying long enough to be studied/identified/traced. [ ADVISORY ] Problem 1: ManTrap hides processes from the caged intruder (or atleast tries to), by hiding them from the /proc in the cage (resembles what many LKM trojans do), which sounds fine. BUT, the cage still runs the SAME kernel as the real box, and also has access to all kernel memory etc. So what we can do is to use a syscall, in this case kill(), which gets its information from the kernel instead of /proc, and compare it's results with the information in /proc. What we do is to simply send a signal (SIGCONT in this case) which hopefully shouldn't affect anything, to PID 1 to 65535 as a nonroot user, and compare the results with /proc. kill() gives EPERM and /proc/ exists -> Fine kill() gives EPERM and /proc/ does not exist -> Not fine! This can also possibly be used to detect LKM trojanss and the like. It might give a false alarm though, as some kernel patches designed to hide other user's processes might give the same result. But together with the other tell-tale signs of ManTrap it gives a very good fingerprint. Problem 2&3: This looks like a result from the /proc filtering/emulation mentioned above. /proc/.. doesn't show up in any syscalls wanting to get the directory listing! (such as getdents()). Also, (cd /proc/self; cd cwd; pwd) gives an error. Another interesting note is that the whole box can be made to lock up and disconnect all users for a couple of minutes by doing: # cd /proc && cd self && cd cwd # pwd # cd ../../../../../ # cd proc # cd self # ls, pwd, etc, Problem 4: ManTrap accidentally seems to hide the first 4 processes always running on a Solaris box (sched, init, fsflush, pageout). Problem 5: Since / isn't the real root of the filesystem, but /usr/rti/cage/root, the inode number is very off, usually in the 100000-200000 range instead of the normal 2. This can be checked by simply doing `ls -id /`. Problem 6: If the intruder got root in the cage, it's very possible to read/write directly to/from /dev/mem, the raw disk device[s], etc. The crash(1M) utility can be used to examine /dev/mem and get the real process listing etc, which includes all ManTrap's processes. (Yes, they can be killed...) More serious damage can be caused using the raw disk device[s], such as /dev/rdsk/c0t0d0s0. ANY data on the system can be read/modified by the intruder, which can be used to wipe logs and such. An utility such as fsdb(1M) can be used to view the directory listings etc. [ EXPLOIT? ] We've written a small program to demonstrate problems 1-3. It's capable of identifying ManTrap using these methods, and also list all 'hidden' processes. It's available from . [ FINAL WORDS ] This basically shows that you can't rely upon anything but a total instruction-level emulation to make a real-looking and yet secure cage. We look forward to such a product as it would be a great tool in intrusion detection. As VMware shows, this can be done atleast on x86 CPUs and it would surprise me if it wouldn't be possible on other platforms (such as Sparc). One key thing to take in mind when deploying a honeypot is to take special care in the architecture of the deployment. Place particular attention on where in the network it is deployed. It is reccomended that if an organization does deploy any honeypots, the systems are firewalled off in a secured, second DMZ. ================================================================ Loki Fate Research Labs loki@f8labs.com ---------------------------------------------------------------- BEGIN PGP SIGNATURE iQA/AwUBOfZvfGnwBJRV5bxfEQJu7gCfQ/T0O9u75nzRGWVSeurNmnFRVr8Anj0c M+UXhPDBvsm+ffRpv41zevQN =3IRx ================================================================