Aum Shinrikyo and the AI Uplift Problem: How LLMs Could Lower the Threshold for CBRN Harm

Abstract

This case study examines how large language models could lower the barriers to chemical, biological, radiological, nuclear, and explosives harm by improving access to technical knowledge, troubleshooting, interpretation, and operational decision-making. Using Aum Shinrikyo as a historical case, it argues that AI risk evaluations should measure not only whether models provide dangerous information, but whether they help capable actors overcome the practical bottlenecks that previously limited their success.