Date: November 20, 2026
Time: 3-4 p.m.
Speaker: Alaia Solko-Breslin, fifth-year Ph.D. student, University of Pennsylvania
Title: Efficient and Sound Probabilistic Verification for AI Agents
Abstract: AI agents increasingly interact with complex digital environments, making it important to ensure that their actions adhere to security policies. Runtime monitoring approaches that express and enforce policies in formal languages such as Datalog offer a promising solution, but existing approaches are largely restricted to deterministic policies. In practice, however, enforcing security policies often requires reasoning about uncertain outcomes. For example, declassifiers and Personally Identifiable Information (PII) detectors may fail with some probability on each invocation, introducing probabilistic predicates into the verification problem.
Furthermore, the outcomes of these events may be correlated, making the independence assumptions required by existing probabilistic inference approaches unsuitable. We address this problem with a sound and efficient verification framework based on distributionally robust optimization. Our approach computes sound upper bounds on the probability of policy violation without making assumptions about correlations between probabilistic predicates. On standard benchmarks involving terminal and tool-calling agents, we demonstrate that our approach improves the security-utility trade-off over prior approaches while providing rigorous guarantees on the probability of policy violation.
Bio: Alaia Solko-Breslin is a fifth-year Ph.D. student at the University of Pennsylvania, where she is advised by Rajeev Alur, and a Student Researcher at Google DeepMind. Her research centers around using formal methods to make machine learning more trustworthy, with a recent focus on agent safety and security.