The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands using ffmpeg and mkdir in Step 3 to process video frames. The variable INPUT, representing a user-supplied file path or URL, is directly interpolated into the command strings (e.g., ffmpeg -i INPUT ...). This pattern is vulnerable to command injection if the input is not strictly validated or escaped to prevent the execution of arbitrary shell commands.
[REMOTE_CODE_EXECUTION]: The ability to construct shell commands with user-controlled variables like INPUT allows for potential unauthorized code execution if an attacker can manipulate the input string to include command separators or subshells.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing external, untrusted media data.
Ingestion points: User-provided video clips and URLs (Stage 0, Step 1, SKILL.md).
Boundary markers: Absent. The skill does not use specific delimiters or instructions to ignore potential malicious content embedded within the media analysis results used in Step 5.
Capability inventory: Shell execution (ffmpeg), media analysis (analyze_media, transcribe_audio), and video generation (generate_reference_video).
Sanitization: None specified. The skill does not describe any filtering or sanitization of data returned from media analysis tools before it is interpolated into the final AI prompt.
Mitigation: A mandatory human-in-the-loop check is implemented in Step 6, requiring the user to approve the final prompt and cost before the generation tool is called.

vfx