The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands for ffmpeg and ffprobe using variables like VIDEO, START, and DURATION. Without proper sanitization, these variables can be used to inject additional shell commands if they contain metacharacters such as backticks or subshell expansions.
[COMMAND_EXECUTION]: AI-generated text is passed directly to the compose_cover.py script and ffmpeg filters. If an attacker influences the AI to output malicious strings, it could lead to exploitation of shell or script parsing.
[PROMPT_INJECTION]: The skill analyzes video frames using AI vision, creating a surface for indirect prompt injection. Attackers could embed instructions in the video to manipulate the AI's generated metadata.
Ingestion points: Video frames processed by media_understand in SKILL.md.
Boundary markers: None identified to separate video content from instructions.
Capability inventory: Capability to execute shell commands (ffmpeg, ffprobe) and local scripts (python3).
Sanitization: Uses double quotes in shell commands, which do not prevent all forms of command injection (e.g., subshells).

video-highlight