simpo-training

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

• Third-party content exposure detected (high risk: 0.85). Training uses dataset_mixer entries like HuggingFaceH4/ultrafeedback_binarized / argilla/... which are public Hugging Face datasets fetched at runtime; their prompt/chosen/rejected text becomes LLM-readable training context (indirect prompt injection risk via dataset content).

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

• Potentially malicious external URL detected (high risk: 0.80). The quick-start installs remote code via "git clone https://github.com/huggingface/alignment-handbook.git" followed by "python -m pip install .", which fetches and executes third-party code as a required setup dependency (remote repository executed locally).