The Agent Skills Directory

[COMMAND_EXECUTION] (HIGH): The skill provides templates for shell commands (e.g., ffmpeg -i input.mp4). If an AI agent populates these placeholders with untrusted user input without strict shell escaping, it enables arbitrary command execution (e.g., providing a filename like ; rm -rf / ;).
[DATA_EXFILTRATION] (MEDIUM): The skill explicitly includes the -safe 0 flag in the concatenation example (ffmpeg -f concat -safe 0 -i filelist.txt). This flag disables security boundaries in the FFmpeg concat demuxer, allowing the tool to read any file on the filesystem that the user has access to. If an attacker can influence the content of the filelist.txt, they can exfiltrate sensitive local data by encoding it into the output video stream.
[REMOTE_CODE_EXECUTION] (HIGH): Because the skill's purpose is to generate executable shell strings from external inputs (Category 8 surface), the lack of sanitization logic or boundary markers constitutes a high-risk RCE vector through Indirect Prompt Injection.
Ingestion points: File paths, codec names, and filter parameters derived from user requests or external file lists (SKILL.md).
Boundary markers: Absent. The skill provides raw shell templates without delimiters or escaping instructions.
Capability inventory: Execution of ffmpeg and ffprobe via the system shell (SKILL.md).
Sanitization: Absent. No guidance is provided to the agent on how to validate or escape inputs before command construction.

ffmpeg-master