evaluation-resource · Documentary review
WMDP
Measures hazardous knowledge in biosecurity, chemical security and cybersecurity.
OriginCenter for AI Safety and the WMDP paper authorsTopicsBio / CBRN · CyberStatusReviewed
Can support
Recognition and selection of answers to WMDP's hazardous-knowledge questions under the stated prompt and model configuration.
Cannot support by itself
Operational CBRN or cyber capability, experimental competence, tool use, planning, intent, tacit knowledge, end-to-end harm creation, or deployment risk.
Decision use
Best used for
Testing hazardous-knowledge retention and evaluating some unlearning methods while monitoring broad capability side effects.
Not enough for
Claims that a model can execute a biological, chemical, or cyber operation or that removing benchmark knowledge eliminates misuse risk.