New arXiv paper proposes framework to measure AI's CBRN risk uplift

A research team led by Rahul Gupta and 10 co-authors from institutions including Boston University and Sandia National Laboratories has released a 19-page preprint on arXiv proposing the Threshold Exceedance Criteria (TEC) framework for evaluating Chemical, Biological, Radiological, and Nuclear (CBRN) uplift in frontier language models. The framework decomposes uplift studies into independent components: participant eligibility, threat scope definition, and statistical estimation of material uplift. The study distinguished between generative uplift (model assisting plan creation from scratch) and revisionist uplift (model assisting refinement of an existing plan).

The empirical study, conducted under controlled pre-release evaluation, produced attack plans across all four CBRN domains evaluated by subject-matter experts. Key findings revealed domain heterogeneity: model-assisted plans sometimes received expert-equivalent instructional ratings, but confirmed material uplift was limited to the radiological domain. The authors emphasize that these findings informed mitigation and deployment-governance decisions rather than characterizing deployed model behavior. They conclude with methodological lessons including prespecified criteria, explicit baselines, and careful distinction between preliminary screening signals and confirmed risk determinations.