The Agent Skills Directory

Command Execution (HIGH): The scripts/download_video.sh script is vulnerable to shell command injection. The $QUERY variable is interpolated within a double-quoted string ("ytsearch${COUNT}:${QUERY}"), allowing subshell execution (e.g., $(id)) if the query is derived from untrusted input.\n- Dynamic Execution (HIGH): scripts/generate_tts.sh uses a shell HEREDOC to construct and execute a Python script. Shell variables like $VOICE are interpolated into the Python source code by the shell before execution, leading to both shell command injection and Python code injection risks.\n- Command Execution (MEDIUM): scripts/render_all.sh executes a Python snippet via python3 -c that interpolates the $SCENES_FILE variable directly into a string. A malicious filename could trigger Python code execution.\n- Indirect Prompt Injection (LOW): The skill is highly vulnerable to indirect injection as it processes scenes.json, which is intended to be generated by an agent.\n
Ingestion points: scenes.json is parsed by multiple scripts to extract IDs, text, and search queries.\n
Boundary markers: None; the scripts treat the JSON content as trusted.\n
Capability inventory: Full system access via ffmpeg, yt-dlp, and npx remotion subprocess calls.\n
Sanitization: No validation or escaping is performed on the data extracted from the JSON before it is used in shell commands.\n- External Downloads (LOW): The skill downloads media from YouTube and other external sources using yt-dlp and curl. While intended, the lack of sanitization on search queries makes these operations dangerous.

search-video-on-web-and-gen