grpo-rl-training

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

• Third-party content exposure detected (high risk: 0.80). The skill's required workflow and templates explicitly load public datasets (e.g., load_dataset('openai/gsm8k') in templates/basic_grpo_training.py and SKILL.md) and instruct the agent to ingest and use those prompt contents for generation and reward computation, meaning untrusted, user-originated web-hosted data from Hugging Face/other public sources can directly influence training behavior.