huggingface-zerogpu

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • External Resource Integration: The skill references official Hugging Face documentation and established GitHub repositories for specialized CUDA dependencies like flash-attention. These references are provided to assist with environment configuration and ensuring compatibility with the ZeroGPU runtime environment.
  • Dynamic Behavior and Serialization: The skill documents platform-specific behaviors including the monkey-patching of the torch library and the use of pickle-based serialization for process isolation. While these involve dynamic code execution and data handling, they are fundamental architectural features of the ZeroGPU platform designed for hardware sharing and security isolation.
  • Input Handling Surface: The skill describes patterns for applications that ingest user-provided data through Gradio interfaces, such as the generate(prompt) example in SKILL.md. This ingestion represents a potential indirect prompt injection surface. The documentation encourages practices like using tempfile for unique output paths to mitigate concurrency issues when processing these inputs.
  • Resource Management and Build Constraints: It provides guidance on managing GPU quotas and durations, as well as addressing build-time constraints like the absence of nvcc in the build phase. This involves recommending the use of pre-built wheels for CUDA-dependent packages.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 04:30 PM
Security Audit — agent-trust-hub — huggingface-zerogpu