The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The script scripts/local_ffmpeg_tool.py contains logic to download a static FFmpeg binary from a third-party domain (https://johnvansickle.com/ffmpeg/releases/ffmpeg-release-amd64-static.tar.xz) when the utility is not found on the local system.
[REMOTE_CODE_EXECUTION]: Once downloaded, the skill extracts the FFmpeg binary, grants it execution permissions via os.chmod, and runs it using the subprocess module. The download implementation in scripts/local_ffmpeg_tool.py includes a bypass that disables SSL certificate validation (ssl.CERT_NONE) if a secure connection cannot be established, which is a significant security flaw.
[COMMAND_EXECUTION]: The tool uses subprocess.run in scripts/main.py and scripts/local_ffmpeg_tool.py to perform media operations. These commands are built by concatenating arguments directly from user-provided inputs (such as source URLs and file paths). While the use of list-based arguments mitigates some risks, it still creates a potential path for command injection if inputs are not strictly validated.
[INDIRECT_PROMPT_INJECTION]: The understand_video_content function accepts user-supplied prompts and URLs to be processed by an AI model, creating a surface for injection attacks.
Ingestion points: CLI parameters parsed in scripts/main.py and passed to the media handling functions.
Boundary markers: The skill lacks explicit separators or instructions to the model to ignore potentially malicious commands embedded in the processed media data.
Capability inventory: The skill possesses capabilities for file system access, network communication, and shell command execution.
Sanitization: There is no evidence of robust sanitization or filtering for natural language prompts before they are interpolated into the model's request context.

byted-mediakit-process-tools