The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied scene descriptions and media files to generate video content. While this creates an attack surface for indirect injection, the implementation follows safe patterns for handling external data.
Ingestion points: Scene descriptions provided via the --subject argument and file paths via --file, --video-file, and --audio-file in scripts/generate-seedance.sh.
Boundary markers: The script does not implement explicit boundary markers or instructions to ignore embedded commands in user input within the director brief construction, relying on downstream model guardrails.
Capability inventory: The script performs file uploads to a remote API, executes network requests via curl, and downloads/opens resulting media files using the open command on macOS.
Sanitization: User-provided subjects are safely encoded into JSON payloads using Python's json.dumps library in the json_string helper, preventing command injection within the JSON structure.
[EXTERNAL_DOWNLOADS]: The script downloads generated video and image content from the official api.muapi.ai service following successful generation. The downloads are restricted to media file extensions to prevent execution.
[DYNAMIC_EXECUTION]: The shell script utilizes a Python one-liner to ensure safe JSON encoding of user input before transmission to the API. The code is static and does not incorporate untrusted input into the execution string itself.

muapi-seedance-2