18-213/18-613, Summer 2024
Attack Lab: Understanding Buffer Overflow Bugs
1 Introduction
This assignment involves generating a total of five attacks on two programs which have different security vulnerabilities. In this lab, you will:
• Learn different ways that attackers can exploit security vulnerabilities when programs do not safe- guard themselves well enough against buffer overflows.
• Gain a better understanding of how to write programs that are more secure, as well as some of the features provided by compilers and operating systems to make programs less vulnerable.
• Gain a deeper understanding of the stack and parameter-passing mechanisms of x86-64 machine code.
• Gain a deeper understanding of how x86-64 instructions are encoded.
• Gain more experience with debugging tools such as GDB and OBJDUMP.
Note: In this lab, you will gain firsthand experience with methods used to exploit security weaknesses in operating systems and network servers. Our purpose is to help you learn about the runtime operation of programs and to understand the nature of these security weaknesses so that you can avoid them when you write system code. We do not condone the use of any other form of attack to gain unauthorized access to any system resources.
You will want to study Sections 3.10.3 and 3.10.4 of the book as reference material for this lab.
2 Logistics
As usual, this is an individual project. You will generate attacks for target programs that are custom gener- ated for you.
2.1 Getting Files
You can obtain your files from the Autolab site
https://ics.autolabproject.com
After logging in to Autolab, select Attacklab -> Download handout. The Autolab server will build your files and return them to your browser in a tar file called targetk .tar, where k is the unique number of your target programs.
Note: It takes about 15 seconds to build and download your target, so please be patient.
Save the targetk .tar file in a (protected) Andrew directory in which you plan to do your work. Then login to a shark machine and give the command: tar -xvf targetk .tar. This will extract a directory targetk containing the files described below.
You should only download one set of files. If for some reason you download multiple targets, choose one target to work on and delete the rest.
Warning: If you expand your targetk .tar on a PC, by using a utility such as Winzip, or letting your browser do the extraction, you’ll risk resetting permission bits on the executable files.
The files in targetk include:
README .txt: A file describing the contents of the directory
ctarget: An executable program vulnerable to code-injection attacks
rtarget: An executable program vulnerable to return-oriented-programming attacks
cookie .txt: An 8-digit hex code that you will use as a unique identifier in your attacks.
farm .c: The source code of your target’s “gadget farm,” which you will use in generating return-oriented programming attacks.
hex2raw: A utility to generate attack strings.
In the following instructions, we will assume that you have already copied the files to a protected local directory, and that you are executing the programs in that local directory.
2.2 Important Points
Here is a summary of some important rules regarding valid solutions for this lab. These points will not make much sense when you read this document for the first time. They are presented here as a central reference of rules once you get started.
• You must do the assignment on one of the class shark machines. There are ten machines available to students to use, a full list of which can be found on the course Web site at:
http://www.cs.cmu.edu/ ˜18213/labmachines.html
• Your solutions may not use attacks to circumvent the validation code in the programs.
• You may only construct gadgets from filertarget with addresses ranging between those for func- tions start_farm and end_farm.
• You can use any gadgets you discover in the farm, not just those for which we give the bytecodes in Figure 3.
• You are allowed to use the standard tools for this assignment: gdb (or lldb), objdump, and any tool that you create (from scratch) on your own.
• You are not allowed to use tools or gdb plugins that are designed to assist in implementing buffer overflow and ROP attacks.
• It is highly recommended that you gdb to confirm that your exploit is placed on the stack with the correct byte ordering
• It is not uncommon for gdb to get lost while performing exploit code. You may be unable to step through your exploitingdb while seeing all of the usual diagnostic information.
3 Target Programs
Both CTARGET and RTARGET read strings from standard input. They do so with the function read and process line defined below:
1 / * read_and_process_line - Read a line from INFILE and process it .
2 This function has a buffer overflow vulnerability . */
3 void read_and_process_line(FILE *infile) {
4 unsigned char buf[BUFFER_SIZE];
5 unsigned char *sp = buf;
6 int c;
7 while ((c = getc(infile)) != EOF && c != ’\n’) { 8 *sp++ = c;
9 }
10 process_line(buf, sp - buf);
11 }
This function reads a byte sequence from standard input, terminated by either a newline (’\n’), or end of file (EOF). Then it calls another function, process line, passing it the bytes that were read. (It does not add a NUL terminator to the bytes it reads, so what it passes to process line is not a “string.” Instead, it passes the number of bytes read as process line’s second argument.)
In the code sample, you can see that read and process line stores the byte sequence in a local vari- able buf, an array of BUFFER_SIZE bytes. (BUFFER_SIZE is a compile-time constant, specific to your version of CTARGET and RTARGET.) Notice that the while loop does not stop when BUFFER_SIZE bytes have been read. This is the same bug that’s found in the C library function gets: it keeps reading data until end of line or file,possibly overrunning the bounds of the storage allocated for the data.
If the input read by read and process line is sufficiently short, nothing interesting will happen:
$ . /ctarget
Cookie: 0x599051eb
Type string: Keep it short!
No exploit, read_and_process_line returned normally.
Typically an error occurs if you type a long string:
$unix ./ctarget
Cookie: 0x599051eb
Type string: This is not a very interesting string, but it is quite long Ouch!: You caused a segmentation fault!
Better luck next time
(Note that the value on the Cookie: line will differ from yours.)
of the cookie shown will differ from yours.) Program RTARGET will have the same behavior. As the error message indicates, overrunning the buffer typically causes the program state to be corrupted, leading to a memory access error. Your task is to be more clever with the strings you feed CTARGET and RTARGET so that they do more interesting things. These are called exploit strings.
Both CTARGET and RTARGET take several different command line arguments: -h: Print list of possible command line arguments
-i FILE: Supply input from a file, rather than from standard input
Your exploit strings will typically contain byte values that do not correspond to the ASCII values for printing characters. The program HEX2RAW will enable you to generate these raw strings. See Appendix A for more information on how to use HEX2RAW.
Important points:
• Your exploit string must not contain byte value 0x0a at any intermediate position, since this is the ASCII code for newline (‘\n’). When Gets encounters this byte, it will assume you intended to terminate the string.
• HEX2RAW expects two-digit hex values separated by one or more white spaces. So if you want to create a byte with a hex value of 0, you need to write it as 00. To create the word 0xdeadbeef you should pass “ef be ad de” to HEX2RAW (note the reversal required for little-endian byte ordering).
When you have correctly solved one of the levels, your target program will automatically send a notification to Autolab. For example:
|
Phase
|
Program
|
Level
|
Method
|
Function
|
Points
|
|
1
|
CTARGET
|
1
|
CI
|
touch1
|
10
|
|
2
|
CTARGET
|
2
|
CI
|
touch2
|
25
|
|
3
|
CTARGET
|
3
|
CI
|
touch3
|
25
|
|
4
5
|
RTARGET RTARGET
|
2
3
|
ROP ROP
|
touch2 touch3
|
35
5
|
CI: Code injection
ROP: Return-oriented programming
Figure 1: Summary of attack lab phases
$unix ./hex2raw < ctarget .l2 . txt | ./ctarget
Cookie: 0x1a7dd803
Type string:Touch2!: You called touch2(0x1a7dd803) Valid solution for level 2 with target ctarget
PASSED: Sent exploit string to server to be validated . NICE JOB!
Unlike the Bomb Lab, there is no penalty for making mistakes in this lab. Feel free to fire away at CTARGET and RTARGET with any strings you like.1
Figure 1 summarizes the five phases of the lab. As can be seen, the first three involve code-injection (CI) attacks on CTARGET, while the last two involve return-oriented-programming (ROP) attacks on RTARGET.
4 Part I: Code Injection Attacks
For the first three phases, your exploit strings will attack CTARGET. This program is set up in a way that the stack positions will be consistent from one run to the next and so that data on the stack can be treated as executable code. These features make the program vulnerable to attacks where the exploit strings contain the byte encodings of executable code.
4.1 Level 1
For Phase 1, you will not inject new code. Instead, your exploit string will redirect the program to execute an existing procedure.
Function read and process line is called within CTARGET by a function test having the following C code:
1 void test(FILE *infile) {
2 read_and_process_line(infile);
3 notify_fail(0, "No exploit, read_and_process_line returned normally . ");
4 }
When read and process line executes its return statement (line 5 of read and process line), the program ordinarily resumes execution within function test (at line 5 of this function). We want to change this behavior. Within the file ctarget, there is code for a function touch1 having the following C representation:
1 TOUCH_FN touch1(void) {
2 vlevel = 1; / * Part of validation protocol */
3 printf("Touch1!: You called touch1()\n");
4 validate(1); 5 }
Your task is to get CTARGET to execute the code for touch1 when read and process line executes its return statement, rather than returning to test. Note that your exploit string may also corrupt parts of the stack not directly related to this stage, but this will not cause a problem, since touch1 causes the program to exit directly.
Some Advice:
• All the information you need to devise your exploit string for this level can be determined by exam- ining a disassembled version of CTARGET. Use objdump -d to get this dissembled version.
• The idea is to position a byte representation of the starting address for touch1 so that the ret instruction at the end of the code for read and process line will transfer control to touch1.
• Be careful about byte ordering. Remember, arrays (such as strings) are saved in index order, but values like integers are evaluated in little-endian.
• You might want to use GDB to step the program through the last few instructions of read and process line to make sure it is doing the right thing.
• The placement of buf within the stack frame. for read and process line depends on the value of compile-time constant BUFFER_SIZE, as well the allocation strategy used by GCC. You will need to examine the disassembled code to determine its position.
4.2 Level 2
Phase 2 involves injecting a small amount of code as part of your exploit string.
Within the file ctarget there is code for a function touch2 having the following C representation:
1 TOUCH_FN touch2(unsigned val) {
2 vlevel = 2; / * Part of validation protocol */
3 if (val == cookie) {
4 printf("Touch2!: You called touch2(0x% . 8x)\n", val);
5 validate(2); 6 } else {
7 notify_fail(2, "Misfire: You called touch2(0x% . 8x)", val);
8 }
9 }
Your task is to get CTARGET to execute the code for touch2 rather than returning to test. In this case, however, you must make it appear to touch2 as if you have passed your cookie as its argument.
Some Advice:
• You will want to position a byte representation of the address of your injected code in such a way that ret instruction at the end of the code for read and process line will transfer control to it.
• Recall that the first argument to a function is passed in register %rdi.
• Your injected code should set the register to your cookie, and then use a ret instruction to transfer control to the first instruction in touch2.
• Do not attempt to use jmp or call instructions in your exploit code. The encodings of destination addresses for these instructions are difficult to formulate. Use ret instructions for all transfers of control, even when you are not returning from a call.
• See the discussion in Appendix B on how to use tools to generate the byte-level representations of instruction sequences.
4.3 Level 3
Phase 3 also involves a code injection attack, but passing a string as argument.
Within the file ctarget there is code for functions hexmatch and touch3 having the following C representations:
1 / * Compare string to hex represention of unsigned value */
2 static int hexmatch(unsigned val, char *sval) {
3 char *endp;
4 unsigned long cval = strtoul(sval, &endp, 16);
5 return (cval == (unsigned long) val 6 && endp != sval
7 && *endp == ’\0’);
8 } 9
10 TOUCH_FN touch3(char *sval) {
11 vlevel = 3; / * Part of validation protocol */
12 if (hexmatch(cookie, sval)) {
13 report_touch3("Touch3!", sval);
14 validate(3);
15 } else {
16 report_touch3("Misfire", sval);
17 notify_fail(3, "touch3 called with the wrong cookie"); 18 }
19 }
Your task is to get CTARGET to execute the code for touch3 rather than returning to test. You must make it appear to touch3 as if you have passed a string representation of your cookie as its argument.
Some Advice:
• You will need to include a string representation of your cookie in your exploit string. The string should consist of the eight hexadecimal digits (ordered from most to least significant) without a leading “ 0x.”
• Recall that a string is represented in C as a sequence of bytes followed by a byte with value 0. Type “man ascii” on any Linux machine to see the byte representations of the characters you need.
• Your injected code should set register %rdi to the address of this string.
• When functions hexmatch and strncmp are called, they push data onto the stack, overwriting portions of memory that held the buffer used by read and process line. As a result, you will need to be careful where you place the string representation of your cookie.
5 Part II: Return-Oriented Programming
Performing code-injection attacks on program RTARGET is much more difficult than it is for CTARGET, because it uses two techniques to thwart such attacks:
• It uses randomization so that the stack positions differ from one run to another. This makes it impos- sible to determine where your injected code will be located.
• It marks the section of memory holding the stack as nonexecutable, so even if you could set the program counter to the start of your injected code, the program would fail with a segmentation fault.
Fortunately, clever people have devised strategies for getting useful things done in a program by executing existing code, rather than injecting new code. The most general form. of this is referred to as return-oriented programming (ROP) [1, 2]. The strategy with ROP is to identify byte sequences within an existing program that consist of one or more instructions followed by the instruction ret. Such a segment is referred to as a gadget. Figure 2 illustrateshow the stack can be set up to execute a sequence of n gadgets. In this figure, the stack contains a sequence of gadget addresses. Each gadget consists of a series of instruction bytes, with the final one being 0xc3, encoding theret instruction. When the program executes a ret instruction starting with this configuration, it will initiate a chain of gadget executions, with the ret instruction at the end of each gadget causing the program to jump to the beginning of the next.
A gadget can make use of code corresponding to assembly-language statements generated by the compiler, especially ones at the ends of functions. In practice, there may be some useful gadgets of this form, but not enough to implement many important operations. For example, it is highly unlikely that a compiled function
Figure 2: Setting up sequence of gadgets for execution. Byte value 0xc3 encodes theret instruction.
would have popq %rdi as its last instruction before ret. Fortunately, with a byte-oriented instruction set, such as x86-64, a gadget can often be found by extracting patterns from other parts of the instruction byte sequence.
For example, one version of rtarget contains code generated for the following C function:
void setval_210(unsigned *p)
{
*p = 3347663060U;
}
The chances of this function being useful for attacking a system seem pretty slim. But, the disassembled machine code for this function shows an interesting byte sequence:
0000000000400f15 <setval 210>:
|
400f15: c7 07 d4 48 89 c7
|
movl
|
$0xc78948d4,(%rdi)
|
|
400f1b: c3
|
retq
|
|
The byte sequence 48 89 c7 encodes the instruction movq %rax, %rdi. (See Figure 3A for the encodings of useful movq instructions.) This sequence is followed by byte value c3, which encodes the ret instruction. The function starts at address 0x400f15, and the sequence starts on the fourth byte of the function. Thus, this code contains a gadget, having a starting address of 0x400f18, that will copy the 64-bit value in register %rax to register %rdi.
Your code for RTARGET contains a number of functions similar to the setval_210 function shown above in a region we refer to as the gadget farm. Your job will be to identify useful gadgets in the gadget farm and use these to perform attacks similar to those you did in Phases 2 and 3.
Important: The gadget farm is demarcated by functions start_farm and end_farm in your copy of rtarget. Do not attempt to construct gadgets from other portions of the program code.
A. Encodings of movq instructions
movq S , D
Figure 3: Byte encodings of instructions. All values are shown in hexadecimal.
5.1 Level 2
For Phase 4, you will repeat the attack of Phase 2, but do so on program RTARGET using gadgets from your gadget farm. You can construct your solution using gadgets consisting of the following instruction types, and using only the first eight x86-64 registers (%rax–%rdi).
movq : The codes for these are shown in Figure 3A.
popq : The codes for these are shown in Figure 3B.
ret : This instruction is encoded by the single byte 0xc3.
nop : This instruction (pronounced “no op,” which is short for “no operation”) is encoded by the single
byte 0x90. Its only effect is to cause the program counter to be incremented by 1. Some Advice:
• All the gadgets you need can be found in the region of the code for rtarget demarcated by the functions start_farm and mid_farm.
• You can do this attack with just two gadgets.
• When a gadget uses a popq instruction, it will pop data from the stack. As a result, your exploit string will contain a combination of gadget addresses and data.
5.2 Level 3
Before you take on the Phase 5, pause to consider what you have accomplished so far. In Phases 2 and 3, you caused a program to execute machine code of your own design. If CTARGET had been a network server, you could have injected your own code into a distant machine. In Phase 4, you circumvented two of the main devices modern systems use to thwart buffer overflow attacks. Although you did not inject your own code, you were able inject a type of program that operates by stitching together sequences of existing code. You have also gotten 95/100 points for the lab. That’s a good score. If you have other pressing obligations consider stopping right now.
Phase 5 requires you to do an ROP attack on RTARGET to invoke function touch3 with a pointer to a string representation of your cookie. That may not seem significantly more difficult than using an ROP attack to invoke touch2, except that we have made it so. Moreover, Phase 5 counts for only 5 points, which is not a true measure of the effort it will require. Think of it as more an extra credit problem for those who want to go beyond the normal expectations for the course.
To solve Phase 5, you can use gadgets in the region of the code in rtarget demarcated by functions start_farm and end_farm. In addition to the gadgets used in Phase 4, this expanded farm includes the encodings of different movl instructions, as shown in Figure 3C. The byte sequences in this part of the farm also contain 2-byte instructions that serve as functional nops, i.e., they do not change any register or memory values. These include instructions, shown in Figure 3D, such as andb %al,%al, that operate on the low-order bytes of some of the registers but do not change their values.
Some Advice:
• You’ll want to review the effect a movl instruction has on the upper 4 bytes of a register, as is described on page 183 of the text.
• The official solution requires a sequence of eight gadgets. Depending on the contents of your farm (each target has a different one), you may be able find a shorter one.
• Remember: Your exploit string must not contain the newline character (byte value 0x0a) at any inter- mediate position