WMDP

Name: WMDP
Creator: Center for AI Safety & collaborators
License: https://spdx.org/licenses/MIT.html

by Center for AI Safety & collaborators

What is WMDP? WMDP is a safety & factuality eval made by Center for AI Safety & collaborators. 3,668 multiple-choice questions that act as a stand-in for hazardous knowledge in biosecurity, cybersecurity and chemical security. It is used both to measure how much risky knowledge a model holds and to test unlearning methods that try to remove it.

Page last updated 2026-10-18 · Evals Review

Editor note

3,668 multiple-choice questions that act as a stand-in for hazardous knowledge in biosecurity, cybersecurity and chemical security. It is used both to measure how much risky knowledge a model holds and to test unlearning methods that try to remove it.

What it tests: Hazardous dual-use knowledge and how well unlearning removes it.
Original source: www.wmdp.ai/
Last updated: Repo last commit 2025-05-29
License: MIT

Facts come only from the linked source (checked 15–16 Oct 2026). Credit belongs to the creators. We don't score or rank evals: how we review.

Reviews

No reviews yet.

Reviews are written by people who sign in with GitHub, one per account per eval. We can't confirm that each reviewer used the eval; we remove reviews that break our rules, never just for being negative, and we never add, buy or edit reviews. Averages, when shown, come only from these reviews.

Reviews from practitioners who sign in with GitHub open soon.

Related safety & factuality evals

HarmBench — Robustness to jailbreaks and harmful requests.
TruthfulQA — Truthfulness versus imitating human misconceptions.
StrongREJECT — Real effectiveness of jailbreak attacks.
SimpleQA — Short-form factuality and calibration.

All Safety & factuality evals