Join us live on Zoom >
( Register to receive a reminder )
Wednesday, Mar 04, 2026 04:30pm - 05:30pm ET

Discovering and Controlling AI Safety Risks in Foundation Models: A Probabilistic Perspective

Mar 04, 2026

Abstract

As foundation models, including large language models and multimodal models, are increasingly deployed in complex and high-stakes settings, ensuring their safety has become more important than ever. In this talk, I present a probabilistic perspective on AI safety: safety risks are treated as structured distributions to be discovered and controlled, rather than isolated failures to be patched. I first introduce probabilistic red-teaming methods that characterize distributions of failures, revealing systematic safety risks that standard evaluation often misses. I then describe probabilistic defense methods that control model behavior during deployment by adaptively steering generation toward constraint-aligned distributions. By unifying failure discovery and behavior control under a probabilistic perspective, this talk highlights a distributional approach for understanding and managing safety risks in foundation models.