Definition
Plain language
A benchmark testing whether AI models have hazardous knowledge that could enable mass-harm uses.
As stated in the literature
The Weapons of Mass Destruction Proxy benchmark, designed to probe hazardous knowledge in biology, chemistry, and cybersecurity domains in LLMs.
Also called: W-M-D-P
Why it matters: It lets labs measure and target the specific knowledge they want to keep out of public models, supporting unlearning and safety evaluations.
For example, WMDP might ask a model technical questions about synthesizing dangerous pathogens to see whether it answers them.
Heard on the show
“One of them is biosecurity — specifically the WMDP benchmark, which probes hazardous biology knowledge.”Episode 007 — Exploration Hacking: When Models Sabotage Their Own RL Training