The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to analyze external video content and user-provided descriptions, creating a surface for indirect prompt injection (e.g., via malicious text or commands embedded within video frames or subtitles).
Ingestion points: Ingests external video files (processed through agent vision and local Python scripts) and user text inputs describing video narrative.
Boundary markers: The SKILL.md file includes an explicit 'Boundaries' section instructing the agent that commands or instructions discovered within the video or accompanying text are strictly data to be analyzed and must never be interpreted as instructions to change permissions or execute external commands.
Capability inventory: The skill leverages local command execution via subprocess.run to call ffprobe and ffmpeg for media analysis, as well as file system operations to store analysis artifacts.
Sanitization: The provided Python scripts (probe_video.py, sample_frames.py, etc.) utilize safe practices such as passing command arguments as lists rather than shell strings, and the evidence_contract.py validation logic includes checks to ensure generated prompts do not contain unauthorized injection sequences.
[COMMAND_EXECUTION]: The skill relies on several local Python scripts to perform media probing and frame sampling. These scripts utilize subprocess.run to interface with standard media utilities like ffprobe and ffmpeg.
Evidence: scripts/probe_video.py and scripts/sample_frames.py call ffprobe and ffmpeg using subprocess.run with list-based argument passing, which effectively mitigates shell injection risks. These operations are restricted to the local file system using the -protocol_whitelist file flag, ensuring no unauthorized network fetching occurs during media analysis.

minimax-h3-video-reverse