It's possible to enumerate which Windows API calls are hooked by an EDR using inline patching technique, where a jmp
instruction is inserted at the beginning of the syscall stub to be hooked.
{% page-ref page="../code-injection-process-injection/how-to-hook-windows-api-using-c++.md" %}
{% page-ref page="bypassing-cylance-and-other-avs-edrs-by-unhooking-windows-apis.md" %}
{% page-ref page="../code-injection-process-injection/api-monitoring-and-hooking-for-offensive-tooling.md" %}
Below shows the stub for for NtReadVirtualMemory
on a system with no EDR present, meaning the syscall NtReadVirtualMemory
is not hooked:
We can see the NtReadVirtualMemory
syscall stub starts with instructions:
00007ffc`d6dcc780 4c8bd1 mov r10,rcx
00007ffc`d6dcc783 b83f000000 mov eax,3Fh
...
{% hint style="info" %}
The above applies to most routines starting with Zw
, i.e ZwReadVirtualMemory
too.
{% endhint %}
...which translates to the following 4 opcodes:
4c 8b d1 b8
4c 8b d1 b8
- are important for this lab - we will come back to this in a moment in a section Checking for Hooks.
Below shows an example of how NtReadVirtualMemory
syscall stub looks like when it's hooked by an EDR:
Note that in this case, the first instruction is a jmp
instruction, redirecting the code execution somewhere else (another module in the process's memory):
jmp 0000000047980084
...which translates to the following 5 opcodes:
e9 0f 64 f8 c7
{% hint style="info" %}
e9
- opcode for near jump
0f64f8c7
- offset, which is relative to the address of the current instruction, where the code will jump to
{% endhint %}
Knowing that interesting functions/syscalls (that are often used in malware), starting with Nt
| Zw
, before hooking, start with opcodes: 4c 8b d1 b8
, we can determine if a given function is hooked or not by following this process:
- Iterate through all the exported functions of the ntdll.dll
- Read the first 4 bytes of the the syscall stub and check if they start with
4c 8b d1 b8
- If yes, the function is not hooked
- If no, the function is most likely hooked (with a couple of exceptions mentioned in the False Positives callout).
Below is a simplified visual example attempting to further explaine the above process:
NtReadVirtualMemory
starts with opcodese9 0f 64 f8
rather than4c 8b d1 b8
, meaning it's most likely hookedNtWriteVirtualMemory
starts with opcodes4c 8b d1 b8
, meaning it has not been hooked
{% hint style="warning" %}
False Positives
Although highly effective at detecting functions hooked with inline patching, this method returns a few false positives when enumerating hooked functions inside ntdll.dll, such as:
`NtGetTickCount
NtQuerySystemTime
NtdllDefWindowProc_A
NtdllDefWindowProc_W
NtdllDialogWndProc_A
NtdllDialogWndProc_W
ZwQuerySystemTime`
The above functions are not hooked. {% endhint %}
Below is the code that we can compile and run on an endpoint running an AV/EDR to see enumerate APIs that were most likely hooked:
#include <iostream>
#include <Windows.h>
int main()
{
PDWORD functionAddress = (PDWORD)0;
// Get ntdll base address
HMODULE libraryBase = LoadLibraryA("ntdll");
PIMAGE_DOS_HEADER dosHeader = (PIMAGE_DOS_HEADER)libraryBase;
PIMAGE_NT_HEADERS imageNTHeaders = (PIMAGE_NT_HEADERS)((DWORD_PTR)libraryBase + dosHeader->e_lfanew);
// Locate export address table
DWORD_PTR exportDirectoryRVA = imageNTHeaders->OptionalHeader.DataDirectory[IMAGE_DIRECTORY_ENTRY_EXPORT].VirtualAddress;
PIMAGE_EXPORT_DIRECTORY imageExportDirectory = (PIMAGE_EXPORT_DIRECTORY)((DWORD_PTR)libraryBase + exportDirectoryRVA);
// Offsets to list of exported functions and their names
PDWORD addresOfFunctionsRVA = (PDWORD)((DWORD_PTR)libraryBase + imageExportDirectory->AddressOfFunctions);
PDWORD addressOfNamesRVA = (PDWORD)((DWORD_PTR)libraryBase + imageExportDirectory->AddressOfNames);
PWORD addressOfNameOrdinalsRVA = (PWORD)((DWORD_PTR)libraryBase + imageExportDirectory->AddressOfNameOrdinals);
// Iterate through exported functions of ntdll
for (DWORD i = 0; i < imageExportDirectory->NumberOfNames; i++)
{
// Resolve exported function name
DWORD functionNameRVA = addressOfNamesRVA[i];
DWORD_PTR functionNameVA = (DWORD_PTR)libraryBase + functionNameRVA;
char* functionName = (char*)functionNameVA;
// Resolve exported function address
DWORD_PTR functionAddressRVA = 0;
functionAddressRVA = addresOfFunctionsRVA[addressOfNameOrdinalsRVA[i]];
functionAddress = (PDWORD)((DWORD_PTR)libraryBase + functionAddressRVA);
// Syscall stubs start with these bytes
char syscallPrologue[4] = { 0x4c, 0x8b, 0xd1, 0xb8 };
// Only interested in Nt|Zw functions
if (strncmp(functionName, (char*)"Nt", 2) == 0 || strncmp(functionName, (char*)"Zw", 2) == 0)
{
// Check if the first 4 instructions of the exported function are the same as the sycall's prologue
if (memcmp(functionAddress, syscallPrologue, 4) != 0) {
printf("Potentially hooked: %s : %p\n", functionName, functionAddress);
}
}
}
return 0;
}
Below is a snippet of the output of the program compiled from the above source code and run on a system with an EDR present. It shows some of the interesting functions (not all displayed) that are most likely hooked, with an exception of NtGetTickCount
, which is a false positive, as mentioned earlier:
After I've posted this note on my twitter, I got the following reply from Derek Rynd:
Derek is suggesting to check if the syscall
instruction itself is not hooked. The syscall
handler routine (responsible for locating functions in the SSDT based on a syscall number) location can be found by reading the Model Specific Register (MSR) at location 0xc0000082
and confirming that the address stored there points to nt!KiSystemCall64Shadow
.
Below shows how this could be done manually in WinBDG:
lkd> rdmsr c0000082
msr[c0000082] = fffff803`24a13180
lkd> u fffff803`24a13180
nt!KiSystemCall64Shadow:
fffff803`24a13180 0f01f8 swapgs
fffff803`24a13183 654889242510900000 mov qword ptr gs:[9010h],rsp
{% embed url="https://posts.specterops.io/adventures-in-dynamic-evasion-1fe0bac57aa" %}
{% embed url="https://rayanfam.com/topics/hypervisor-from-scratch-part-8/" %}