Yue Zhao

Research Summary

My research focuses on AI auditing, safety, and security for agent systems, with related work on foundation models and their deployment in real-world environments. I study how modern AI systems fail, how their behavior and risks can be evaluated, monitored, and controlled at the system level, and how they can be deployed responsibly in domains where failures carry significant consequences. This agenda is organized around three closely connected directions:

1. AI Auditing & Assurance
I develop methods, benchmarks, and open-source systems for auditing foundation models and agent systems. This work includes trustworthiness evaluation, security analysis of AI pipelines, ecosystem-scale risk scanning, and continuous monitoring that provides evidence for assurance in deployment. Representative systems include agent-audit and TrustLLM.

Keywords: AI Auditing, AI Assurance, Trustworthy AI, Foundation Models, Agent Systems, TrustLLM, agent-audit, Monitoring, Evaluation Frameworks, Risk Analysis

2. AI Safety & Security
I study failure modes, attack surfaces, and runtime control layers in large language models and agent systems, and design methods to detect and mitigate unsafe or compromised behavior. Representative topics include hallucinations, jailbreaks, prompt attacks, privacy leakage, model extraction, failures in multi-agent interactions, and runtime guardrails such as policy enforcement and approval workflows for tool-using agents. Representative systems include Aegis. This direction is also informed by my earlier work on anomaly detection and out-of-distribution detection.

Keywords: LLM Safety, Agent Safety, AI Security, AI Agent Security, Runtime Guardrails, Policy Enforcement, Hallucination Mitigation, Jailbreak Detection, Prompt Attacks, Privacy Leakage, Model Extraction, Robustness, OOD Detection, Anomaly Detection

3. AI for Science & Society
I apply reliable and auditable AI systems to domains where correctness, safety, and accountability matter, including climate and weather forecasting, healthcare and biomedicine, and computational social systems. These applications also serve as demanding testbeds for auditing, assurance, and safety methods.

Keywords: AI for Science, Climate AI, Weather Forecasting, Healthcare AI, Biomedicine, Computational Social Systems, Decision Modeling, High-Stakes AI