The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The scripts/convert_to_gguf.py script clones the llama.cpp repository from GitHub (https://github.com/ggerganov/llama.cpp.git) and executes its internal Python scripts and compiled binaries. This pattern of downloading and running code from non-trusted third-party sources is a significant risk vector.\n- [COMMAND_EXECUTION]: Multiple scripts, particularly scripts/convert_to_gguf.py, use subprocess.run to execute shell commands for git operations, build processes (cmake, make), and package management (pip). The use of shell=True was not detected, but the execution of downloaded binaries poses a risk.\n- [PRIVILEGE_ESCALATION]: Instructions in references/gguf_conversion.md and logic in scripts/convert_to_gguf.py recommend or attempt to perform system package installations using apt-get, which requires root/sudo privileges.\n- [INDIRECT_PROMPT_INJECTION]: The skill's core purpose is to ingest datasets from the Hugging Face Hub (via load_dataset) for training and preference alignment (SFT, DPO, GRPO). This creates a vector for indirect prompt injection if the dataset contents are malicious and intended to influence the model's behavior or escape the training environment.\n- [DYNAMIC_EXECUTION]: The skill utilizes the hf_jobs MCP tool to submit and run dynamically generated Python script strings on remote GPU infrastructure. It also performs runtime compilation and execution of the llama-quantize binary.\n- [EXTERNAL_DOWNLOADS]: The skill downloads external dependencies and scripts from GitHub and Hugging Face. While Hugging Face is a trusted organization, the skill also targets the ggerganov/llama.cpp repository, which is a third-party resource.

huggingface-llm-trainer