WMDP is a benchmark of 4,157 multiple-choice questions surrounding hazardous knowledge in biosecurity, cybersecurity, and chemical security, aimed to measure and mitigate the risk of LLM misuse. Towards reducing these risks, we introduce CUT, an unlearning method which reduces LLM performance on WMDP while retaining performance on general tasks.

HarmBench is an evaluation framework for automated language model red teaming. We conduct a large-scale comparison of 18 red teaming methods and 33 target LLMs and defenses, proposing a highly efficient adversarial training method that enhances LLM robustness.

Representation engineering (RepE) enhances LLM transparency by monitoring and manipulating high-level cognitive phenomena. RepE is effective in mitigating dishonesty, hallucination, and other unsafe behaviors.

MACHIAVELLI is a benchmark of 134 text-based choose-your-own-adventure games with annotations of safety-relevant concepts such as deception, physical harm, and power-seeking, guiding development towards safe yet capable language agents.