“The Weapons of Mass Destruction Proxy (WMDP) benchmark is designed to evaluate unlearning methods intended to remove hazardous knowledge from large language models.”