yt-dlp
does not limit the extensions of downloaded files, which could lead to arbitrary filenames being created in the download folder (and path traversal on Windows). Since yt-dlp
also reads config from the working directory (and on Windows executables will be executed from the yt-dlp directory) this could lead to arbitrary code being executed.
yt-dlp
version 2024.07.01 fixes this issue by whitelisting the allowed extensions.
This means some very uncommon extensions might not get downloaded; however, it will also limit the possible exploitation surface.
It is recommended to upgrade yt-dlp to version 2024.07.01 as soon as possible, always have .%(ext)s
at the end of the output template, and make sure you trust the websites that you are downloading from. Also, make sure to never download to a directory within PATH or other sensitive locations like your user directory, system32
, or other binaries locations.
For users not able to upgrade:
- Make sure the extension of the media to download is a common video/audio/sub/... one
- Try to avoid the generic extractor (--ies default,-generic
)
- Keep the default output template (-o "%(title)s [%(id)s].%(ext)s
)
- Omit any of the subtitle options (--write-subs
, --write-auto-subs
, --all-subs
, --write-srt
)
- Use --ignore-config --config-location ...
to not load config from common locations
One potential exploitation might look like this:
From a mimetype we do not know, we default to trimming the leading bit and using the remainder. Given a webpage that contains
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "VideoObject",
"name": "ffmpeg",
"encodingFormat": "video/exe",
"contentUrl": "https://example.com/video.mp4"
}
</script>
this will try and download a file called ffmpeg.exe
(-o "%(title)s.%(ext)s
).
ffmpeg.exe
will be searched for in the current directory, and so upon the next run arbitrary code can be executed.
Alternatively, when engineering a file called yt-dlp.conf
to be created, the config file could contain --exec ...
and so would also execute arbitrary code.
A big thanks to @JarLob for independently finding a new application of the same underlying issue. More can be read about on the dedicated GitHub Security Lab disclosure here: Path traversal saving subtitles (GHSL-2024-090)
{ "nvd_published_at": "2024-07-02T14:15:13Z", "cwe_ids": [ "CWE-434", "CWE-669" ], "severity": "HIGH", "github_reviewed": true, "github_reviewed_at": "2024-07-02T15:58:35Z" }