The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to execute ffmpeg for extracting audio from video files. This is a standard and safe practice as it uses a list of arguments to prevent command injection.
[EXTERNAL_DOWNLOADS]: The skill guides users to download large model weights (approx. 30GB) from Hugging Face repositories including Wan-AI, TencentGameMate, MeiGen-AI, and hexgrad. These are well-known sources for AI model distribution.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text for TTS (Text-to-Speech) and media files (images/audio/video). While this represents a theoretical ingestion point for malicious content, it is restricted to the primary function of media generation and does not have capabilities for downstream privilege escalation or data exfiltration.
Ingestion points: Text inputs via --text argument and media files via --input_path and --audio_path in scripts/infer_infinitetalk.py.
Boundary markers: None explicitly present in the wrapper script.
Capability inventory: Uses ffmpeg for media conversion and standard AI libraries (torch, transformers, diffusers) for inference.
Sanitization: Input text is processed by standard NLP tokenizers; media files are processed by established libraries like OpenCV, Librosa, and FFmpeg.

infinitetalk