The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/extract_frames.py invokes external binaries ffmpeg and ffprobe to process video files. While it uses argument lists to avoid shell injection, it dynamically constructs complex filter strings for ffmpeg. The --timestamps feature interpolates the video's filename directly into the drawtext filter without escaping single quotes or colons. A maliciously crafted filename (e.g., containing ') could break the filter's syntax, allowing for filter graph manipulation within the ffmpeg process.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted video files and feed the resulting frames to an LLM, creating a vulnerability to indirect prompt injection where visual content (text or images) contains instructions intended to manipulate the agent. * Ingestion points: Video files provided by the user are processed by scripts/extract_frames.py as described in SKILL.md. * Boundary markers: Absent. The instructions do not advise the agent to treat extracted visual content as potentially adversarial. * Capability inventory: The skill can read from the local filesystem, write images to a temporary directory, and execute ffmpeg subprocesses. * Sanitization: Absent. The script extracts and scales frames without verifying the safety of the content within them.

video-frames