It is known that dumping Windows credentials is a technique often utilized for everyday attacks by adversaries and, consequently, Red Teamers. This process has been out there for several years and is well documented by MITRE under the T1003.001 technique. Sometimes, when conducting a Red Team engagement, there may be some limitations when trying to go beyond the early detection of this technique to allow defenders to train complex manipulation and usage of the credentials.
One of the options to overcome this limitation is to explicitly allow the execution of this technique. However, there is another way, which is both stealthier and more lightweight. The following article will dive into how it can be executed.
ReactOS, is an interesting and valuable project for anyone interested in understanding the low-level code of a Windows-like OS. We found that starting with the less-resistant path and trying to compile minidump.c from ReactOS to be quite difficult. However, after carefully analyzing the minidump module from skelsec, we found information about the minidump file format.
The minidump format is quite complex and has many structures, pointers, and sections. In order to keep things as simple as possible, we experimented with the minidump python module to remove and change several parts in order to understand if these were relevant.
A minidump is composed of multiple "streams" which are like sections that contain specific information. For example, ExceptionStream presumably contains information like the stack trace in case the minidump was created due to a crash.
After testing with pypykatz we found that the only relevant streams were SystemInfoStream, ModuleListStream, and Memory64ListStream. This first finding simplified the process because limiting the number of streams reduced the processing that needed to be done.
This stream has information about the Windows machine, and it is not related to LSASS itself. It has relevant information such as the Windows version and build number, but also less relevant fields such as the number of processors.
We ended up setting all the fields that were not needed to NULL. This made the process of creating the minidump a lot simpler, as we were able to ignore irrelevant fields.
All of the DLLs LSASS loaded are listed in this stream. It is worth noting that, while this stream is important, for this exercise, it wasn’t necessary to include every single DLL.
In fact, we were able to ignore most of them and kept only those that are relevant to mimikatz, such as kerberos.dll and wdigest.dll. This decision effectively made the size of the dump a lot smaller.
The actual memory pages of the LSASS process can be found in this stream. However, it takes up a lot of space, so reducing its size was critical to reduce the overall dump size. We decided to ignore any page that met any of the following conditions:
Page wasn’t committed
Page marked as mapped
Page protection equals PAGE_NOACCESS
Page marked as PAGE_GUARD
Ignoring all these pages did not break the analysis of mimikatz, but did effectively reduce the size of the dump.
By taking out all the non-vital information from the dump we managed to reduce the dump from roughly 50MB down to 10MB.
As explained earlier, another goal was to achieve some level of obfuscation. Given that the creation of the minidump is done programmatically, we had full control of the dump and thus could implement any obfuscation that we chose.
We opted to corrupt the "magic bytes" (or signature) of the minidump file format, which is a simple, yet effective approach.
Minidumps start with the string "PMDM" in big endian. Changing these magic bytes would make it more difficult to figure out if a block of memory is a minidump, and since this is at the very start of the file, the binary blob wouldn’t look like a minidump, not even at creation time.
This modification did break mimikatz and pypykatz. We created a small bash post-dump script to restore the original format once the dump is on the tester’s machine.
PID of LSASS
To dump LSASS, you typically need to know the PID of the LSASS process. The action of listing all the running processes could be seen as an abnormal or suspicious activity. Running tasklist or even calling CreateToolhelp32Snapshot might be detected by advance security solutions.
We decided to use the NtGetNextProcess syscall to loop over all the processes in the system until we found a process that had 'lsass.exe' loaded. This was a valid method to find the LSASS process and avoided having to go through the usual steps.
Avoiding API calls
Reducing the number of API calls was important for obvious reasons: userland hooks. The only Windows API call that nanodump calls is LookupPrivilegeValueW, which is used to enable SeDebugPrivilege. This privilege should already be enabled in most cases, but feel free to remove this call if you want to be even stealthier. Besides that, everything is done using syscalls to avoid userland hooks.
To use syscalls, we used SysWhispers2 so, there was no need to re-compile nanodump for every new version of Windows. We had to make a few changes to the code to avoid using global variables given that Beacon Object Files (BOF) do not support them. We also used InlineWhispers to build nanodump on Linux using Mingw.
We also wanted to have the possibility of downloading the dump using Beacon's C2 channel without touching the disk. However, it can be written to a file if need be.
No Beacon? No Problem
As explained earlier, we initially started this project as part of our Red Team practice, allowing us to conduct complex threat actions. Sometimes we don’t need to go as far as deploying Beacon on each compromised machine, so we added the possibility to use the .EXE version of nanodump. The one limitation that exists for the EXE version is that you cannot use the fileless download feature, given that it relies on Cobalt Strike’s C2 channel for it.
While it was challenging creating a SYSCALL based minidump, it was also critical for many scenarios. Additionally, creating a malleable module capable of feeding the great mimikatz is a powerful and flexible approach. The idea of modularizing a software solution has been out there for many years and this context is even more important to improve the success and future updates facing strong and dynamic detection tools.
Do it Yourself
If you’re interested in using nanodump, we’ve posted the code to our Github.
Skelsec for his amazing work with minidump and pypykatz.
freefirex from CS-Situational-Awareness-BOF at Trustedsec for many cool tricks for BOFs
jthuraisamy for SysWhispers2