Misinfo-Jailbreak Bench

Misinfo-Jailbreak Bench is a comprehensive benchmark dataset designed to evaluate the risk of Large Language Models (LLMs) generating structured misinformation under jailbreak attacks. It contains 31,964 prompts, created by systematically applying 11 distinct jailbreak strategies to misinformation generation tasks in both English and Chinese. The benchmark's key innovation is its framework for quantitative, claim-level severity assessment, moving beyond simple refusal metrics. It is a vital resource for diagnosing model vulnerabilities and developing robust safety defenses against adversarial information generation.