Summerschool Aachen 2004/Programming Errors Lab
- 1 Presentation Summary
- 2 Notes from the Lab Session
- 3 Notes from Presentations
Notes from the Lab Session
Solution for case 1a
- use gdb to get the address of buf
- print 20 no-ops (0x90) into a file
- append the shellcode (e.g. from packetstorm) to that file
- fill that file to 104 bytes
- append the address of buf to that file
- run ./case1a "`cat $file`" and voila!
Notes from Presentations
Writing a tunnelling program in order to tunnel arbitrary prtocols into HTTP
We are currently working on a program to tunnel arbitrary protocols into HTTP. The programing language we are using is C using among other libraries the socket.h library.
Exploiting buffer overflows
I worked on stack based buffer overflows. The first part of the task, i.e. just creating a buffer overflow, was quite simple. The size of the buffer was 100 bytes and as 4 bytes are used in order to store the value of the saved frame pointer in the the stack I first thought that an input string of length 105 would already be enough to cause a buffer overflow. However, since gcc allocates more than 100 bytes for the buffer in the stack, 124 bytes of imput were necessary in order to cause a buffer overflow. The second part was more difficut as here, I needed to insert shellcode in order to start a shell AND to overwrite the return address saved in the stack with the address of the shellcode which is supposed to start a shell. In order to accomplish this task I used the tutorial "Smashing the stack for fun and profit", which was very useful.
Metasploit is a nice repository for shellcode, UNIX as well as Windows.
Example vulnerability on the SRCF
As I think I mentioned, the SRCF has a big problem allowing so many users to run CGI and PHP scripts on our webserver. We try to limit the number of insecure scripts by looking for old versions of vulnerable software, but we do not have the resources to check whether user-written scripts are exploitable. What we try to do is stop one user's script from interfering with other user's scripts, and stop attackers from being able to escalate a exploitable CGI into a root exploit.
This approach does not always work and there have been two successful root compromises since 2000. The first was simply getting a user shell using an exploitable CGI (Ikonboard). The attacker then used a ptrace exploit to get root. This was before a patch to the vulnerability has been released. They then installed the SuKIT rootkit.
The more interesting one was where the attacker used site-specific knowledge to get root. It is likely he used a vulnerable CGI to look around the system. What he noticed is that one other vulnerable CGI was written by an ex-sysadmin and so could get access to that account. While that user did not have root access, he did own one old file which was processed by a cron job. This cron job ran as root and did not check for shell metacharacters in the input file (since it was meant to only be written to by sysadmins). The attacker embedded a command in the input file so as to create a root shell when it was processed by the cron job. Fortunately we stopped tha attack before any harm could be done, and we also captured a list of commands he entered. In reaction to this attack we introduced two new policies. Firstly all scripts run as root were audited to ensure that no files writable by non-root users were opened. Secondly each system administrator had a new account created, which was isolated from their normal account, and that could not run CGIs.
Exploiting Buffer Overflows
When I tried to exploit buffer overflows like case 1 on my Fedora Core 2 system I noticed some difficulties. First of all, the place for the local variables on the stack is significantly larger than it was declared in the source file. This is a problem if you want to exploit off-by-one programming errors, because you are not able to overwrite something usefull. The second thing I noticed was even worser. Normally, all memory offsets of local stack variables stay at the same place, even if you restart the program. Of course it depends on the input, but I used the same arguments and the offsets were allways different. It was the Gnu C Compiler that troubled my day. Since version 3.0 Gnu changed the memory alignment of the compiler, but i don't know the details. Because Sami had still the older GCC 2.9 on his system, we started trying to exploit the buffer overflows on his system. We were succesfull with Case 1 and tried Case 2 until it was too late to think any further. Yes, days are getting shorter ;)
--Feanor 22:34, 22 Sep 2004 (CEST)
In the lab session I had to find out that my old FreeBSD shell code didn't work
anymore, because I loaded %al with the syscall number, but missed to initialize
%ah to zero before. They changed the syscall behaviour some time ago, which I
had not realized. This took me some time...
I solved case 1a and used a small perl script for actually exploiting it. Now I'm still working on the other cases. Btw, Ilja has done a very cool preparation. Thanks! 8)
--Cpunkt 11:45, 23 Sep 2004 (CEST)
Wait till the fun stuff arrives.
I researched the php bugs demonstrated at the lecture (originally found by Martin Eizsner), and made some interesting discoveries. Due to security reasons these can't be disclosed yet (they *should* get disclosed once the bugs are fixed).
Then I decided to audit tor and if I found a trivial bug, I could leave these notes at the system running tor. Unfortunatly I didn't find such a trivial bug. I did find a few other issues, which I can't disclose yet either (they should also get disclosed some time soon).
UPDATE: The tor bugs are fixed, you can read about them here: - http://archives.seul.org/or/dev/Oct-2004/msg00003.html - http://archives.seul.org/or/cvs/Oct-2004/msg00032.html
-- Ilja van Sprundel