llama-cpp

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill instructs users to download source code from the official llama.cpp GitHub repository and models from HuggingFace. These are standard, well-known resources for LLM development.
  • Evidence: git clone https://github.com/ggerganov/llama.cpp in SKILL.md.
  • Evidence: huggingface-cli download TheBloke/Llama-2-7B-Chat-GGUF in quantization.md.
  • [REMOTE_CODE_EXECUTION]: The skill describes a workflow where source code is cloned from a remote repository and compiled locally. This is standard for the documented tool and targets a trusted open-source project.
  • Evidence: git clone and make instructions in SKILL.md.
  • [COMMAND_EXECUTION]: The documentation includes multiple examples of command-line operations for building the project and executing inference binaries.
  • Evidence: compilation commands like make LLAMA_METAL=1 and binary execution like ./llama-cli in SKILL.md and optimization.md.
  • [INDIRECT_PROMPT_INJECTION]: As a tool for LLM inference, it possesses an inherent attack surface where the engine processes untrusted user prompts.
  • Ingestion points: User prompts provided to llama-cli in SKILL.md.
  • Boundary markers: Absent in the basic usage examples provided.
  • Capability inventory: The tool facilitates model file reading and network server hosting as documented in SKILL.md and server.md.
  • Sanitization: Absent in the instructional examples.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:08 PM
Security Audit — agent-trust-hub — llama-cpp