Anti-unpacker tricks - part one

2008-12-01

Peter Ferrie

Microsoft, USA
Editor: Helen Martin

Abstract

Unpackers have been around for as long as packers themselves, but anti-unpacking tricks are a more recent development and have grown quickly both in number and, in some cases, complexity. Peter Ferrie describes a range of anti-unpacking tricks.


Unpackers have been around for as long as packers themselves, but anti-unpacking tricks are a more recent development. Anti-unpacking tricks have grown quickly both in number and, in some cases, complexity. This paper is an addendum to a paper presented at the CARO workshop in May this year [1], and describes some of the anti-unpacking tricks that have come to light since that paper was published.

Introduction

Anti-unpacking tricks come in different forms, depending on what kind of unpacker they are intended to attack. Unpackers can be in the form of memory-dumpers, debuggers, or emulators:

  • A memory-dumper dumps the process memory of the running process without regard to the code inside it.

  • A debugger attaches to the process, allowing single-stepping, or the placing of breakpoints at key locations, in order to stop execution at the right place. The process can then be dumped with more precision than a memory-dumper alone.

  • An emulator, as referred to within this paper, is a purely software-based environment, most commonly used by anti- malware software. It places the file to execute inside the environment and watches the execution for particular events of interest.

There are corresponding tricks for each of the above, and they will be discussed separately.

Anti-dumping

Self-unmapping

The data that fills the image space of a process is simply a mapped view of the file. This view can be unmapped using the kernel32 UnmapViewOfFile() function. If the data is moved first to another location, and then any absolute references are adjusted according to the new image base value, then execution can be resumed from the other location. The result is a process that cannot be dumped by ordinary means.

Example code looks like this:

 push 0
 call GetModuleHandleA
 mov ebx, [eax+3ch] ;lfanew
 ;SizeOfImage
 mov ebx, [eax+ebx+50h]
 push 40h ;PAGE_EXECUTE_READWRITE
 push 1000h ;MEM_COMMIT
 push ebx
 push 0
 xchg esi, eax
 call VirtualAlloc
 mov ecx, ebx
 0D 
 lea edi, [eax+offset l1]
 sub edi, esi
 push esi
 push edi
 xchg edi, eax
 rep movsb
 jmp UnmapViewOfFile
l1: ;execution continues here
 ;but in relocated region
 ...

Page redirection

Page redirection is the extreme implementation of the Guard Pages technique that was described in [1], as it applies to Armadillo. In the Guard Pages technique, a guard page is used to allow per-page decryption. It could be defeated by touching the pages, one at a time, and then writing those pages to disk, one at a time. Page redirection avoids that weakness by not restoring the pages to their original location. Instead, all accesses are redirected to other locations in memory where the pages now exist. The result is that the kernel32 ReadProcessMemory() function cannot be used to dump the memory remotely, and the kernel WriteFile() function cannot be used to dump the memory locally using the original addresses, because the redirection will not occur. However, there are two methods that can be used to dump the memory. One is to find the addresses of the redirected pages. This is difficult to automate, and there may be further obfuscation of the content, which will interfere with this method. The second method is simply to perform a user-mode copy of the data, using the original addresses, copying it to a dynamically allocated block of memory. The data can then be written directly from that block of memory. This technique is used by BASLR.

Anti-debugging

Special APIs

NtYieldExecution

The ntdll NtYieldExecution() function is used to allow the currently running thread to give up the rest of its execution period and allow any other scheduled thread to execute. The function returns an error if no threads are scheduled to execute. When an application is being debugged, the act of single-stepping through the code causes debug events, and as a result the debugger thread is always scheduled to resume execution. This fact can also be used to infer the presence of a debugger, though it may also detect the presence of a thread that is running with high priority.

Example code looks like this:

 push 20h
 pop ebp
l1: push 0fh
 call Sleep
 call NtYieldExecution
 cmp al, 1
 adc ebx, ebx
 dec ebp
 jne l1
 inc ebx
 je being_debugged

This technique is used by the Extreme Debugger Detector.

NtSetLdtEntries

Perhaps because the local descriptor table (LDT) is not used by Windows, it is generally not supported properly (or at all) by debuggers. As a result, it can be used as a simple anti-debugger technique. Specifically, a new LDT entry can be created, which maps to some code. Then, by performing a far transfer of control (call or jump) to the new LDT entry, the debugger might become lost or refuse to go further.

Example code looks like this:

 ;base must be <= PE->ImageBase
 ;but no need for 64kb align
 base equ 12345678h
 ;sel must have bit 2 set
 ;CPU will set bits 0 and 1
 ;even if we don’t do it
 sel equ 777h

 xor eax, eax
 push eax
 push eax
 push eax
 ;4k granular, 32-bit
 ;present, DPL3, exec-only code
 ;limit must not touch kernel mem
 ;calculate carefully to use APIs
 push (base and 0ff000000h) \
   + 0c1f800h \
   + ((base shr 10h) and 0ffh)
 push (base shl 10h) + 0ffffh
 push sel
 call NtSetLdtEntries
 ;jmp far sel:l1
 db 0eah
 dd offset l1 – base
 dw sel
l1: ;execution continues here
 ;but using LDT selector
 ...

Turbo Debug32 fails to disassemble the code inside the LDT range, but execution continues correctly. OllyDbg refuses to continue execution within the LDT range. WinDbg disassembles the code correctly inside the LDT range, and execution continues correctly. This technique is used by some malware. It was probably based on some inaccurate documentation on the ReactOS site [2], which misplaces the System bit and includes too many bits in the Type field.

NtQueryInformationProcess

The ntdll NtQueryInformationProcess() function has the following parameters: HANDLE ProcessHandle, PROCESSINFOCLASS ProcessInformationClass, PVOID ProcessInformation, ULONG ProcessInformationLength and PULONG ReturnLength. Windows Vista supports 45 classes of ProcessInformationClass information (up from 38 classes in Windows XP), but so far only four of them have been documented by Microsoft. One of these is the ProcessDebugPort. It is possible to query for the existence (but not the value) of the port. The return value is 0xffffffff if the process is being debugged. Internally, the function queries for the non-zero state of the EPROCESS->DebugPort field. A common method for hiding the debugger process from the ProcessDebugPort class is to zero the value, but without checking the process handle for which the presence of the port is being queried. This presents a problem when a debugger is supposed to be present (for example, in Armadillo), since a debug port should exist for that process. This problem has been disclosed publicly [3].

Example code looks like this:

 xor ebx, ebx
 mov ebp, offset l1
 push ebp
 call GetStartupInfoA
 ;sizeof(PROCESS_INFORMATION)
 sub esp, 10h
 push esp
 push ebp
 =0 D 
 push ebx
 push ebx
 push 1 ;DEBUG_PROCESS
 push ebx
 push ebx
 push ebx
 push ebx
 push offset l2
 call CreateProcessA
 pop eax
 push eax
 mov ecx, esp
 push 0
 push 4 ;ProcessInformationLength
 push ecx
 push 7 ;ProcessDebugPort
 push eax
 call NtQueryInformationProcess
 pop eax
 test eax, eax
 je being_faked
 ...
 ;sizeof(STARTUPINFO)
l1: db 44h dup (?)
l2: db "myfile", 0

CloseHandle

As with an invalid handle, if a protected handle is passed to the kernel32 CloseHandle() function (or directly to the ntdll NtClose() function) and no debugger is present, then an error code is returned. However, if a debugger is present, an EXCEPTION_HANDLE_NOT_CLOSABLE (0xC0000235) exception will be raised. This exception can be intercepted by an exception handler, and is an indication that a debugger is running.

Example code looks like this:

 xor eax, eax
 push offset being_debugged
 push d fs:[eax]
 mov fs:[eax], esp
 push eax
 push eax
 push 3 ;OPEN_EXISTING
 push eax
 push eax
 push 80000000h ;GENERIC_READ
 push offset l1
 call CreateFileA
 push eax
 ;HANDLE_FLAG_PROTECT_FROM_CLOSE
 push 2
 push -1
 xchg ebx, eax
 call SetHandleInformation
 push ebx
 call CloseHandle
 ...
l1: db "myfile", 0

Defeating this method is easiest on Windows XP, where a FirstHandler Vectored Exception Handler can be registered by the debugger to hide the exception and silently resume execution. Of course, there is the problem of hooking the kernel32 AddVectoredExceptionHandler() function transparently, in order to prevent another handler from registering as the first handler. However, it is still easier than transparently hooking the ntdll NtClose() function on Windows NT and Windows 2000 in order to register a Structured Exception Handler to hide the exception. This method has been disclosed publicly [4].

NtSystemDebugControl

The ntdll NtSystemDebugControl() function could have been a very nice function for detecting debuggers. It was introduced in Windows NT, and its capabilities increased in subsequent versions of Windows. It supported a SysDbgQueryModuleInformation command, which was an alternative to the SystemProcessInformation class of the ntdll NtQuerySystemInformation() function. Windows XP introduced the SysDbgReadVirtual command, which allowed the reading of virtual memory from anywhere in the system. There were other commands for writing to virtual memory, reading and writing physical memory and MSRs, among others. Alas, in Windows 2003 SP1 and later, all of these functions were disabled. The functions that remain are for enabling and disabling the kernel debugger, and querying and setting some minor behaviours.

Non-continuable exceptions

When an exception occurs, the flags specify whether it is continuable or not. A continuable exception is one for which the cause can be corrected, a nd then execution can resume from the location at which the exception occurred. An example of a continuable exception is a memory access violation. The use of such an exception is how user-mode paging is implemented by packers such as Shrinker.

A non-continuable exception is one for which, under normal circumstances, the cause cannot be corrected. An example of a non-continuable exception is a division by zero. Any attempt to resume execution after a non-continuable exception is raised will cause Windows to issue an EXCEPTION_INVALID_DISPOSITION exception, prior to terminating the application. However, through the use of a breakpoint in the ntdll RtlRaiseException() function, it is possible to interfere with that sequence. Specifically, the context that is saved on the stack prior to the breakpoint exception can be altered to clear the non-continuable flag. Once that is done, execution of the ntdll RtlRaiseException() function can be resumed. At that point, the call to ntdll RtlRaiseException() becomes indistinguishable from an ordinary call. EXCEPTION_INVALID_DISPOSITION is still delivered to the exception handler, but the application is no longer terminated upon return from the handler. Further, the cause can be corrected. In the case of a division by zero, the divisor can be replaced with a non-zero value, and then the division can be attempted again. This technique has been disclosed publicly [5].

Hardware tricks

Besides the prefetch queue trick that was described in [1], there is another trick that detects single-stepping. It has worked since the earliest of Intel CPUs, and was common in DOS, but it still works in all versions of Windows. The trick relies on the fact that certain instructions cause all interrupts to be disabled for one instruction. In particular, loading the SS register clears interrupts for one instruction in order to allow the [E]SP register to be used without risk of stack corruption. Of course, the [E]SP register does not need to be loaded. Any instruction can follow the load of the SS register. If a debugger is being used to single-step through the code, then the T flag will be set in the EFLAGS image. This is typically not visible because a debugger will clear the image whenever the flags are saved. However, if the debugger cannot gain control in time to intercept the save, then it has no way of hiding the T flag. Specifically, the debugger cannot gain control if all interrupts are disabled.

Example code looks like this:

push ss
pop ss
pushfd
test b [esp+1], 1
jne being_debugged

This technique is used by the20ChupaChu debugger test.

Device names

Tools that make use of kernel-mode drivers also need a way of communicating with those drivers. A very common method is through the use of named devices. Any success when attempting to open such a device indicates the presence of the driver.

Example code looks like this:

 xor eax, eax
 mov edi, offset l2
l1: push eax
 push eax
 push 3 ;OPEN_EXISTING
 push eax
 push eax
 push eax
 push edi
 call CreateFileA
 inc eax
 jne being_debugged
 or ecx, -1
 repne scasb
 cmp [edi], al
 jne l1
 ...
l2: <array of ASCIIZ strings,
 null to end>

Recent lists include the following names:

  • \\.\SPCommand

  • \\.\Syser

  • \\.\S yserBoot

  • \\.\SyserLanguage

  • \\.\SyserDbgMsg

These names belong to Syser.

OllyDbg-specific

A potential arbitrary-code-execution vulnerability was reported recently for OllyDbg [6]. However, the author of this paper discovered that the report is incorrect with respect to the target. The problem is not in OllyDbg, but in the dbghelp.dll that OllyDbg carries. This DLL is used by multiple debuggers, including IDA, SoftICE and WinDbg, but not Turbo Debug32 or Immunity Debugger, for example. It is also shipped as part of Windows itself. The version of the DLL that OllyDbg carries is indeed vulnerable to a stack buffer overflow. The problem was introduced in Windows XP, and corrected in Windows Server 2003. In Windows 2000, where the file was introduced, a fixed-length copy was performed. Since this was unnecessarily large in most cases, and had the potential to cause crashes because of out-of-bounds reads, it was replaced in Windows XP with a string copy. However, the length of the string was not verified prior to performing the copy, leading to the buffer overflow. The fix was to use a string copy with a specified maximum length.

There is a little-known problem in OllyDbg’s analyser of floating-point instructions, which is caused by having no mask on invalid floating-point operation errors [7]. This allows two special values to cause floating-point errors when converting from double-extended precision to integer. The values are +/- 9.2233720368547758075e18. There is a publicly available demonstration that specifies only the positive value as a candidate, without mentioning that the negative value is also a candidate.

Example code looks like this:

 fld t [offset l1]
 ...
l1: dq -1 
 dw 403dh

This is the public version. The alternative code looks like this:

 fld t [offset l1]
 ...
l1: dq -1 
 dw 0c03dh

SoftICE interrupt 0x2D denial of service

When the kernel32 OutputDebugString() function is called and no user-mode debugger is present, eventually the following code is executed:

mov eax, [ebp+8] ;service (1)
;pointer to message structure
mov ecx, [ebp+0ch]
mov edx, [ebp+10h] ;unused (0)
int 2dh

If SoftICE is installed, then the DbgMsg.sys driver is loaded, even if SoftICE isn’t running. DbgMsg.sys hooks interrupt 0x2D and executes this code when interrupt 0x2D is executed:

 push ecx
 push eax
 call l1
 ...
 ;message structure
l1: mov ecx, [esp+8] 
 mov eax, [eax+4] ;bug here

The read from [eax+4] without checking first if the pointer is valid, leads to a kernel-mode crash (blue screen) if the original ECX register value is invalid.

Example code looks like this:

push 1
pop eax
xor ecx, ecx
int 2dh

This interrupt 0x2D bug has already been published [8], but without=2 0any details about exactly why it happens.

Anti-emulating

Hardware tricks

Exception priority

Given the following code:

 xor eax, eax
 push offset l1
 push d fs:[eax]
 mov fs:[eax], esp
 push -1
 popfd
 ;force an exception to occur
 mov eax, [eax]
 ;ExceptionRecord
l1: mov eax, [esp+4]
 ;ExceptionCode
 mov eax, [eax]
l2: ...

What is the value of EAX when l2 is reached? Perhaps surprisingly, it is not EXCEPTION_ACCESS_VIOLATION (0xC0000005). It is EXCEPTION_SINGLE_STEP (0x8000004). What actually happens is that an EXCEPTION_ACCESS_VIOLATION is raised, but its delivery to the debuggee is delayed. The reason is that the trap flag remains set when the ntdll KiUserExceptionDispatcher() function gains control, and then after the first instruction is executed, the single step exception is raised and delivered to the debuggee. Upon returning from the debu ggee, after dealing with the EXCEPTION_SINGLE_STEP, the EXCEPTION_ACCESS_VIOLATION is delivered to the debuggee. This technique has been disclosed publicly [9].

Software interrupts

Interrupt 0x2E

Interrupt 0x2E is an interface for user-mode code to communicate with native kernel-mode APIs. It was introduced in Windows NT, and continues to be supported for compatibility reasons. The interface accepts a function index in the EAX register. A bounds check is performed against this index, prior to dispatching the request if it is valid. However, if the index is out of bounds, then Windows will return a STATUS_INVALID_SYSTEM_SERVICE (0xC000001C) value in the EAX register. If the index is within bounds, then a bounds check is performed against the parameter pointer in the EDX register. If the parameter pointer is out of bounds, then Windows will return a STATUS_ACCESS_VIOLATION (0xC0000005) value in the EAX register. Some emulators do not support this behaviour.

Example code looks like this:

or eax, -1
cdq
int 2eh
int 2eh
cmp eax, 0c0000005h
jne being_debugged

This technique is used by the ChupaChu debugger test.

Closing remarks

Anti-unpacking tricks continue to be developed because the older ones are constantly being defeated. In part two of this series next month, we will describe some tricks that might become common in the future, along with some countermeasures.

The text of this paper was produced without reference to any Microsoft source code or personnel.

twitter.png
fb.png
linkedin.png
hackernews.png
reddit.png

 

Latest articles:

Nexus Android banking botnet – compromising C&C panels and dissecting mobile AppInjects

Aditya Sood & Rohit Bansal provide details of a security vulnerability in the Nexus Android botnet C&C panel that was exploited to compromise the C&C panel in order to gather threat intelligence, and present a model of mobile AppInjects.

Cryptojacking on the fly: TeamTNT using NVIDIA drivers to mine cryptocurrency

TeamTNT is known for attacking insecure and vulnerable Kubernetes deployments in order to infiltrate organizations’ dedicated environments and transform them into attack launchpads. In this article Aditya Sood presents a new module introduced by…

Collector-stealer: a Russian origin credential and information extractor

Collector-stealer, a piece of malware of Russian origin, is heavily used on the Internet to exfiltrate sensitive data from end-user systems and store it in its C&C panels. In this article, researchers Aditya K Sood and Rohit Chaturvedi present a 360…

Fighting Fire with Fire

In 1989, Joe Wells encountered his first virus: Jerusalem. He disassembled the virus, and from that moment onward, was intrigued by the properties of these small pieces of self-replicating code. Joe Wells was an expert on computer viruses, was partly…

Run your malicious VBA macros anywhere!

Kurt Natvig wanted to understand whether it’s possible to recompile VBA macros to another language, which could then easily be ‘run’ on any gateway, thus revealing a sample’s true nature in a safe manner. In this article he explains how he recompiled…


Bulletin Archive

We have placed cookies on your device in order to improve the functionality of this site, as outlined in our cookies policy. However, you may delete and block all cookies from this site and your use of the site will be unaffected. By continuing to browse this site, you are agreeing to Virus Bulletin's use of data as outlined in our privacy policy.