The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/requirements.txt file references versions of common Python libraries that are not currently available on the official PyPI registry, such as numpy==2.4.6, Pillow==12.2.0, opencv-python==4.13.0.92, and imageio-ffmpeg==0.6.0. This configuration poses a significant dependency confusion risk, where the installation of these 'future' versions could lead to the execution of malicious code if an attacker publishes those specific versions to a public registry.
[COMMAND_EXECUTION]: The script scripts/inspect_video.py executes external system tools including ffmpeg, ffprobe, and MediaInfo using the subprocess module. While it uses list-based arguments to prevent shell injection, the execution of these powerful utilities on user-provided media files represents a significant capability surface that could be leveraged if vulnerabilities exist in the media processing binaries.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external video files, which constitutes a surface for indirect prompt injection and related attacks.
Ingestion points: Video files provided as input to the scripts/inspect_video.py tool.
Boundary markers: The skill does not implement boundary markers or specific safety instructions to isolate the processing of potentially malicious media data.
Capability inventory: The toolchain includes capabilities for shell command execution via subprocess.run, file system writes for extracted frames, and complex image/video processing via OpenCV and Pillow.
Sanitization: There is no evidence of input validation or sanitization for the video stream content, with the skill relying entirely on the robustness of third-party media libraries.

ai-video-output-review