If you discover a security vulnerability in ungguf, please report it by opening a GitHub issue with the security label. For sensitive vulnerabilities, contact the maintainers directly.
ungguf reads and parses binary GGUF files using the gguf Python library. While the library is well-maintained, parsing untrusted binary files always carries inherent risk. Only convert GGUF files from sources you trust.
The vllm_sanity.py inference checker uses trust_remote_code=True when loading models with vLLM. This is required for many HuggingFace models but allows arbitrary code execution during model loading. Only run sanity checks against models from trusted sources.
All conversion and inference runs inside Docker containers with NVIDIA GPU access (--runtime=nvidia). Ensure your Docker installation is properly configured and that GPU access is restricted to trusted users.
ungguf does not verify the integrity or authenticity of GGUF input files. A corrupted or maliciously modified GGUF file could produce incorrect safetensors output. Always verify conversion results using the built-in bit-exact verifier.