LLM jailbreak evaluation benchmark. Includes a dataset of prompts from categories of harmful behavior: