trl-training

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

• Third-party content exposure detected (high risk: 0.85). The skill’s runtime workflow uses Hugging Face --dataset_name / --model_name_or_path to fetch datasets/models from external sources (e.g., trl-lib/* or arbitrary user-supplied Hub repos), and those datasets’ free-form text is ingested into the training pipeline and thus into the LLM context during generation/training.