The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted video files (in.mov, small.mp4) which are processed by a multimodal LLM to determine the storyline and generate captions. This presents an indirect prompt injection surface where visual text or specific patterns in the video could influence the LLM's output. 1. Ingestion points: in.mov, small.mp4 (Stage 1). 2. Boundary markers: Absent. 3. Capability inventory: ffmpeg, gemini-3.5-flash, gemini-3-pro-image. 4. Sanitization: Absent.
[COMMAND_EXECUTION]: The skill makes extensive use of ffmpeg and ffprobe for video manipulation. It also suggests generating and executing a detached Python script to manage parallel image generation requests to bypass shell tool timeouts.
[DATA_EXFILTRATION]: The skill explicitly references the use of proxy.env, which is described as containing Cloudflare gateway credentials required for the Gemini client's operation. Accessing such configuration files represents a data exposure surface.

drawing-video