jailbreakbench

JailbreakBench is an open-source robustness benchmark designed for assessing and advancing the jailbreak capabilities of large language models (LLMs). It offers a comprehensive dataset of 200 distinct behaviors for evaluating the efficacy of jailbreaks and defenses, alongside a leaderboard for tracking algorithm performance. Notable features include the ability to load submitted jailbreak strings, access the JBB-Behaviors datasets, and facilitate red-teaming exercises through a dedicated pipeline.