huggingface-local-models

Pass

Audited by Gen Agent Trust Hub on Aug 7, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill references the download and build process for llama.cpp from its official GitHub repository.
  • Downloads source code from the well-known repository github.com/ggml-org/llama.cpp for local compilation.
  • Provides instructions for installation via standard package managers including Homebrew and Winget.
  • [COMMAND_EXECUTION]: The skill guides the agent to execute shell commands for building software, managing local servers, and processing files.
  • Includes build commands using make with various hardware acceleration flags (Metal, CUDA, ROCm).
  • Instructs the agent to run local serving binaries (llama-cli, llama-server) and Python-based conversion scripts (convert_hf_to_gguf.py).
  • [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by consuming data from external APIs to build shell commands.
  • Ingestion points: Data is retrieved from the Hugging Face Tree API (https://huggingface.co/api/models/<repo>/tree/main) and repo pages (https://huggingface.co/<repo>?local-app=llama.cpp) as documented in references/hub-discovery.md.
  • Boundary markers: None present. The instructions do not explicitly warn the agent to sanitize or validate filenames or quant labels retrieved from the remote API.
  • Capability inventory: The agent has access to shell execution via model serving tools and Python scripts.
  • Sanitization: None observed. The skill assumes the integrity of the filenames and quantization labels provided by the Hugging Face Hub registry.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 7, 2026, 08:25 PM
Security Audit — agent-trust-hub — huggingface-local-models