HarmBench

What it does

HarmBench is an open-source framework (MIT) from the Center for AI Safety and university researchers for evaluating automated red-teaming attacks on language models and how robustly models refuse harmful requests. It provides a set of harmful behaviours, including contextual and multimodal ones, a pipeline that generates attacks and completions, and trained classifiers that judge the results.

Compare with

• vsGarak
• vsART
• vsCleverHans

People who use it also use

Not enough marks yet to tell. When people who use HarmBench mark the rest of their stack, it shows up here.