HarmBench

HarmBench from the Center for AI Safety is an evaluation framework for automated red teaming. It ships several hundred harmful behaviors across categories including cybercrime, chemical and biological, misinformation, harassment and copyright, in text and multimodal forms, along with implementations of many attack methods and a fine tuned classifier that judges whether a model's output actually completed the behavior.

It's the most complete open framework for running attacks, not just a prompt list, and it's the basis of a lot of published safety comparisons.