audio-driven-motion

Pass

Audited by Gen Agent Trust Hub on Nov 1, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Model weights for sound effect synthesis are fetched from Hugging Face using the snapshot_download method in assets/motionfield_nodes/nodes_audio.py. The implementation is secured by a strict allow-list (MOSS_REPO_ALLOWLIST) that limits access to a specific community repository for the MLX framework, preventing the skill from downloading data from unauthorized sources.
  • [COMMAND_EXECUTION]: The custom audio node pack utilizes subprocess.run in assets/motionfield_nodes/nodes_audio.py to handle specialized model inference via a bundled runner script and to process audio files using ffmpeg. These operations are well-scoped; user inputs are strictly validated or cast to numeric types before being passed to the commands, and prompt data is exchanged via standard input rather than shell arguments, effectively mitigating command injection risks.
  • [DYNAMIC_EXECUTION]: Automated discovery and loading of ComfyUI nodes is performed in assets/motionfield_nodes/__init__.py using importlib.import_module. This is a standard and necessary pattern for extending the modular ComfyUI ecosystem and is restricted to imports from sibling files within the skill's own asset directory.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external text prompts and audio tracks to drive generative video models, creating a potential ingestion surface.
  • Ingestion points: Template placeholders ({{PROMPT}}) in workflow files like wan22_s2v.api.json and the prompt input field for the OF_MossSFX node.
  • Boundary markers: Absent; user-provided data is directly interpolated into processing pipelines without specific delimiters.
  • Capability inventory: Communicates with external server endpoints via the HTTP client in scripts/comfy_client.py and writes files to the local output directory via assets/motionfield_nodes/nodes_provenance.py.
  • Sanitization: Absent; the skill relies on the integrated generative models' own safety guardrails and content filters.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 1, 2026, 05:10 PM