huggingface-llm-trainer

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.85). Runtime path: the training job script calls datasets.load_dataset(...) on a user-specified Hub dataset (e.g., trl-lib/Capybara or any --dataset), which ingests outsider-authored dataset text into the LLM training context (and the dataset inspector also fetches example rows via the Datasets Server API and prints them as free text).

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The skill includes examples and scripts that fetch and execute remote code at runtime—e.g., running the dataset inspector via https://huggingface.co/datasets/mcp-tools/skills/raw/main/dataset_inspector.py, submitting TRL example scripts from raw GitHub (e.g. https://raw.githubusercontent.com/huggingface/trl/main/examples/scripts/grpo.py), and the GGUF conversion script clones and builds code from https://github.com/ggerganov/llama.cpp.git—all of which are runtime-fetched resources that directly execute remote code.