Audit-to-Patch Pipelines for Secure LLM Agent Systems

AI agents can fail in repeatable, high-impact ways that only surface after deployment. Yue Zhao from USC’s FORTIS Lab is building an audit-to-patch pipeline that detects and audits security risks in LLM agent code and configuration, then proposes safe, reviewable patches with automated checks – stopping safety failures before they ship and reducing regressions over time.

Biography

How do you catch a security flaw in an AI agent before it ships? Yue Zhao is an assistant professor of computer science at USC and director of the FORTIS Lab (Foundations Of Reliable and Trustworthy Intelligent Systems). His research on reliable and safe AI — spanning LLM and agent safety, robustness, and anomaly detection — powers widely used open-source tools, and his project builds audit-to-patch pipelines that detect and fix high-impact risks in LLM agent systems.