trl-training

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

• Third-party content exposure detected (medium risk: 0.30). In the TRL CLI training workflow (e.g., trl sft/dpo/grpo/rloo/reward in SKILL.md), the agent ingests dataset and model text via user-specified --dataset_name / --model_name_or_path from Hugging Face sources, so an outsider can author poisoned dataset content that the runtime fetches without selecting specific items first.