In GUARD: Role-playing to Generate Natural-language Jailbreakings to Test Guideline Adherence of Large Language Models, researchers propose a novel role-playing system for LLMs to collaborate on creating new jailbreaks. This strategy is employed to efficiently induce LLMs into generating responses that violate ethical guidelines or safety measures, hence testing their adherence to such guidelines.

• The system assigns distinct roles to user LLMs.
• A knowledge graph organizes jailbreak characteristics for easier access.
• GUARD’s role-playing generates guideline-violating responses in LLMs.
• The methodology is tested on three open-source LLMs and a commercial LLM.

This paper unfolds an innovative approach to stress-test LLMs before release, emphasizing the significance of proactive safety measures. Further research can look into refining this methodology and extending it to a wider range of AI applications to bolster overall system resilience.