The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/compose_video.py uploads local image frames to https://catbox.moe/user/api.php using curl. While this is intended to provide public URLs for processing by the fal.ai API, it involves transmitting local project assets to an external third-party service.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute ffmpeg, fc-list, and curl. While arguments are passed as lists, the script dynamically constructs complex ffmpeg filter strings (such as zoompan and drawtext) using data extracted from scenes.json without rigorous validation.
[EXTERNAL_DOWNLOADS]: The skill interacts with the fal.ai API (https://queue.fal.run/fal-ai/wan-i2v) to perform remote video generation tasks. Additionally, the script uses urllib.request to poll for status updates from remote endpoints.
[REMOTE_CODE_EXECUTION]: The documentation in SKILL.md and SKILL.en.md instructs the user to run npx remotion, which involves downloading and executing Node.js code from the npm registry at runtime.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes scenes.json without sanitization.
Ingestion points: scripts/compose_video.py reads scenes.json for narration, text overlays, and composition parameters.
Boundary markers: No delimiters or safety instructions are present to prevent embedded commands in the JSON content from influencing the agent's behavior.
Capability inventory: The script can execute shell commands (ffmpeg, curl) and perform network uploads and API requests.
Sanitization: There is no evidence of filtering or sanitization of the input text used in video overlays or narration strings.

video-editor