README.md

code_runner

Summary

After a PoW check, the server generates a MIPS binary, sends it base64-encoded to us, and then runs it. The binary checks 64 bytes we supply to its stdin, and in case it's happy, it lets us give it some shellcode to run. The catch is that generates binaries are kind of different, and there is a time limit.

Analysis

Even though there is PoW, it's still possible to download a significant amount of samples. I let the script to run in the background, and it ended up collecting about 1000. The main() function is always like this:

int main() {
  gettimeofday(&dt, NULL);
  if (checker()) {
    gettimeofday(&t1, NULL);
    dt.tv_usec = ((t1.tv_sec - dt.tv_sec) * 1000000 + t1.tv_usec) - dt.tv_usec;
    if (dt.tv_usec / 100000 < 13)
      read(0, shellcode, 0x34);
    (*(code *)shellcode)();

This means we have 1.3 seconds to find something that the checker would accept. The first two layers are always the same:

uint checker() {
  puts("Faster > ");
  read(0,input,0x100);
  checker_trampoline(input);

followed by

uint checker_trampoline(uchar *input) {
  return checker0(input);

What follows are 16 chained checkers, each of which inspects 4 bytes:

uint checker0(uchar *input) {
  if ((input[1] == input[3]) && (input[0] == input[2]) &&
      (input[0] == -0x26) && (input[3] == 0x1a))
    return checker1(input + 4);
  else
    return 0;
}

and the final checker is always the same:

uint checker16(uchar *input) {
  return 0xc001babe;
}

Attempt 1: angr

This looks simple enough for angr to solve. Indeed, angr cuts through most checkers like a hot knife through butter. What gives it pause are nonlinear checkers like this one:

uint checker3(uchar *input) {
  x1 = input[2] * input[2] - input[1] * input[1];
  if (x1 < 0) x1 = -x1;
  x2 = input[3] * input[3] - input[0] * input[0];
  if (x2 < 0) x2 = -x2;
  if (x2 <= x1) {
    x1 = input[3] * input[3] - input[2] * input[2];
    if (x1 < 0) x1 = -x1;
    x2 = input[0] * input[0] - input[1] * input[1];
    if (x2 < 0) x2 = -x2;
    if (x1 < x2)
      return checker4(input + 4);
    }
  }
  return 0;
}

The good news is that there is always a solution, where inputs are either 0 or 1, so trying each checker with such an extra constraint first lets us get past those without adding any explicit signatures.

This puts us at 15 seconds under pypy3 though, and despite the presence of some glaring inefficiencies, like starting symbolic execution at the very entry point, it's highly unlikely that we can get to 1.3s. So this approach has to be scrapped.

Attempt 2: grouping and signatures

Since checker_trampoline() is always the same, and each checker follows the same pattern:

if (!cond1) goto fail;
if (!cond2) goto fail;
...
return checker_n+1(input + 4);  # jal ADDR
fail:
return 0;                       # jr ra

we can very quickly identify and extract every single one of them from all samples, and then do a pairwise difflib.SequenceMatcher().ratio() on them in order to find similar groups.

There are certain immediately noticeable thresholds: 98% and 70%. Using 98% produces groups that differ only by constant values (good), but there are way too many of them to further handle manually. Using 70% produces a more manageable amount of groups, however, the differences within each group are quite significant - it's not only register numbers, but sometimes also sequences of instructions.

I'm pretty sure an AV guru could've make it work, but having stared at samples for quite some time, I knew that I would have more luck with the next approach.

Solution: signatures + DIY femtoangr

angr has a lot of nice features, like using a proper IR and forking states, but this challenge doesn't really need them, and they appear slow us down. What we need is converting an essentially linear sequence of checks into a constraint and letting z3 solve it.

There are only ~20 various MIPS instructions in collected samples, so it was fairly easy to write such a convertor. There are a bunch of places where we can cheat and get away with it: not handling overflows, considering all "branch taken" events as lava and matching abs() by signature.

Combined with nearly instantaneous extraction of checkers from the second approach this puts us at 0.2 seconds, which is more than enough.

There are plenty of MIPS shellcodes floating on the internet, I ended up used this one, because it fit somewhat tight space requirements. The combined exploit worked!

Flag

De1ctf{9d94bc3d3f57f1b33aee728b5d32d6d473df8df3}