RTPE: A Prompt Evolution Framework for LLM Red Teaming

RTPE is a prompt evolution framework for LLM red teaming that automatically generates a large number of high-quality, diverse red-team prompts along two dimensions: breadth and depth. Experiments show that RTPE outperforms existing representative automated red-teaming methods in both attack success rate and diversity. Based on 4,800 prompts generated by RTPE, the authors systematically analyzed the performance of 8 representative LLMs on 8 sensitive topics.