The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external media via image_url, video_url, and audio_url fields, which represent a surface for indirect injection attacks where instructions could be embedded in the metadata or content of the media files.
Ingestion points: image_url, video_url, and audio_url fields processed by the runcomfy CLI in SKILL.md.
Boundary markers: The input is wrapped in a JSON structure, and the documentation notes the CLI avoids shell expansion of the prompt content.
Capability inventory: Invokes the runcomfy CLI which performs network requests to the Model API and downloads generated video files to a user-specified directory.
Sanitization: The skill documentation explicitly acknowledges image-based injection risks and implements a 2 GiB download cap to prevent resource exhaustion.
[SAFE]: The skill utilizes official domains (runcomfy.com, runcomfy.net) and a dedicated CLI tool (@runcomfy/cli) associated with the service being integrated.
[SAFE]: Sensitive information is managed securely; the skill instructs users to use environment variables or local configuration files with restricted owner-only permissions (0600).

seedance-v2