huggingface-spaces

Pass

Audited by Gen Agent Trust Hub on Jul 3, 2026

Risk Level: SAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to use the Hugging Face CLI (hf) and pip for environment configuration, repository management, and log monitoring. These commands are standard operations for building and deploying applications on the Hugging Face platform.
  • [EXTERNAL_DOWNLOADS]: Downloads prebuilt CUDA wheels from a dedicated Hugging Face Hub repository (multimodalart/zerogpu-blackwell-wheels) and installs dependencies from reputable research organizations, such as Facebook Research and Inria. These references are used to ensure compatibility with specialized hardware like ZeroGPU.
  • [DYNAMIC_EXECUTION]: Implements runtime package installation using subprocess.run to resolve environment-specific dependency conflicts. It also uses monkey-patching for torch.load and utilizes PyTorch Ahead-of-Time (AoTI) compilation to optimize model loading and execution on ZeroGPU hardware.
  • [INDIRECT_PROMPT_INJECTION]: The skill outlines the creation of Gradio applications that process user-provided prompts. While this establishes a data ingestion surface, the instructions include appropriate patterns for process isolation (ZeroGPU worker forks) and temporary file handling to minimize cross-session risks.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 3, 2026, 01:07 PM
Security Audit — agent-trust-hub — huggingface-spaces