RTPE is a novel evolution framework that automates the discovery of prompts that elicit harmful responses from large language models, enabling more efficient security testing before deployment.

• Creates diverse adversarial prompts through evolutionary algorithms that mutate and recombine successful attacks
• Significantly more efficient than manual red teaming, making safety testing scalable
• Identifies vulnerabilities that might otherwise remain undetected until real-world deployment
• Helps developers proactively strengthen safety guardrails by revealing weak points

This research addresses critical security concerns by providing a systematic approach to identify harmful outputs before LLMs reach users, reducing potential reputation damage and harmful societal impacts.

Be a Multitude to Itself: A Prompt Evolution Framework for Red Teaming