WMDP and the dual-use evaluation stack
WMDP: 4,157 yellow-zone MCQs across three domains
biosecurity
1,520 questions
pathogens, acquisition, synthesis
cybersecurity
2,225 questions
exploit dev, persistence, C2
chemistry
412 questions
schedule-1 precursors
the uplift narrative (2024-2025)
2024: "mild uplift"
novices benefit modestly vs search
April 2025 (OpenAI PF v2): "on the cusp"
cusp of meaningfully helping novices
2025 Anthropic: 2.53x bio-acquisition uplift
insufficient to rule out ASL-3
novice-vs-expert decomposition
novice-relative uplift
multiplicative; high because novices have little
WMDP-shaped + acquisition trial measures this
expert-absolute capability
ceiling of information extractable by motivated expert
elicitation-hardened evaluation measures this