This PoC demonstrates how to infect a running process on a system in order to hide and later trigger code under an attacker's control. An attacker-controlled, parasitic shared library is forced to be loaded into the target process, and a function used normally by the process is redirected to point to a nefarious function living in the parasite library. The basic process is as follows:

1. The parasite library is built by the attacker (in our case libtest.so.1.0) that overrides a function which is known to exist in the address space of the target process (and which is known to be used by the target program). For example, printf is a good bet, since it lives in libc and is pretty commonly used. Our example does this using a function called evilprint(), which prints "I am evil." rather than the original printf() arguments.

2. Once an attacker compromises a system (and gets privilege), this parasite library is brought onto the compromised system, along with an attack tool.

3. The attack tool (in our case p01snr) is run (as root) and uses ptrace to attach to the target process. This will only work as a root user.

4. After attaching, the attack tool injects some shellcode into the target process's code segment (.text) (using ptrace(POKE_TEXT)). The purpose of this shellcode is to open() and subsequently mmap() the shared library into the target process's address space. Newer systems will prevent this sort of code injection, in which case more clever use of ptrace() is necessary. See countermeasures for more detail.

5. The attack tool now uses ptrace() to invoke the mmap() shellcode (using ptrace(PTRACE_CONT)). It then restores the original code in .text that was overwritten in the previous code injection step. At this point the attacker's library is loaded into the target process (which can be verified by inspecting /proc/<PID>/maps of the target process).

6. Now that the library is loaded, the attack tool overrides the original target function (printf()) with the attacker's function inside the parasite library (evilprint()). It does this by patching the GOT entry corresponding to the target function. The attack tool can find the GOT by reading the memory image of the target process at the location where the target program's binary is mapped. This way it can parse the ELF binary and thus find the GOT location. The GOT entry for printf() is overridden with evilprint(), thus redirecting any future calls of the target function to the parasite library. To make this work, the attacker needs to know the address of the overriding function in the parasite library. One way to find this easily is to use a function signature (first few bytes of the object code) to do pattern matching of the library's binary. That's exactly what this PoC does. See scripts/extract_func_sig.sh.

7. The previous will completely override the original target function. One behavior that might be beneficial is to instead orverride it, but then after our overriding function is invoked, invoke the original target function as well. In our case, that would mean that evilprint() does what it needs to, then invokes printf(). This gives us the ability to make it look to the program like nothing has changed. In order for this to occur, the overriding function cannot just directly invoke the original (think about why). Instead, the overriding function contains an obvious call via function pointer to a place-holder address. The attack tool will then play the role of the dynamic linker at runtime and patch this place-holder address with the address of the target function (printf()). It does this by looking for a instruction signature of the obvious call (referred to as "transfer code" within the PoC) and patching it.

That's the basic structure of the attack.

There are a few countermeasures that can prevent this kind of attack from happening:

• Prevent privilege escalation. If an attacker can't get root, they can't ptrace(PTRACE_ATTACH) to a target process, rendering this attack impossible.
• Compile long-running programs statically. In statically compiled binaries, there is no dynamic linking, no GOT, and no shared libraries, so short of rewriting the program's binary, there's no way to perform this attack. Even binary rewriting will likely be impossible due to the next countermeasure. Static linking is not a realistic option for large programs and for heavily used systems, since processes which use the library will duplicate it in memory, wasing sapce. Binary image sizes will skyrocket.
• Use Linux GRSEC patches. These Linux kernel patches prevent (among other things) modifications to a program's .text segment by a tracer using ptrace(). However, we can get around this by avoiding the use of shellcode entirely. We instead write our system call arguments into the .data section of the target, and use ptrace(PTRACE_SYSCALL) to cause the target to stop before the execution of its next system call. We then change the registers (again using ptrace) to invoke a series of different syscalls (in our case open() followed by mmap()) which has the same effect of loading the attack library. The program is the reset to what it was doing before. This work-around is done by our PoC, and you can see it in action in grsec_mmap_library().

There are three components to this PoC. The first is a simple target process (called daemon) which sits in the background and prints a string every few seconds. The second is the parasite library (libtest.so.1.0, implemented in parasite.c) which has an overriding function for printf called evilprint(). Once this library is injected into the target, it will print "I am evil" instead of the original print. The third component is the injection tool, called p01snr. You can learn about its options by running it without arguments (./p01snr). You'll probably want to provide the -g flag to bypass the GRSEC code injection protection (so that our attack will use the method outlined in the previous section). The attack will not likely work without this flag on newer systems.

$ make
$ sudo make install
$ ./daemon & 

The first command builds all the code, in addition to autogenerating the mmap() shellcode payload based on some assembly and autogenerating the function signature string for our attack function. You should take a look at how these are generated by looking at scripts/gen_mmap.sh, scripts/gen_shellcode.sh, and scripts/extract_func_sig.sh.

The second command installs the parasite library into a directory on the default library search path, in this case /lib.

The third command invokes the target as a background process. Once you run this, keep track of the PID that the shell outputs for the daemon. If you lose track of it, you can find it later on with ps aux | grep daemon.

Once we have the PID of the daemon process, we can attack it as follows:

$ sudo ./p01snr -p PID -f puts -l libtest.so.1.0 -g

Note that we must run this command as root, otherwise the attack tool will not be able to ptrace(PTRACE_ATTACH) to the target process.

You should now see the daemon hijacked (evident from it printing "I am evil.").

The PoC code here was originally (way back in 2013) taken from a great article by Ryan (elfmaster) O'Neill. I've since modified it and adapted it for a security class, and added some tools to autogenerate payloads. For more great ELF hackery I recommend perusing his github.