AI systems are deployed faster than they can be verified. Foundation models and autonomous agents now make consequential decisions, execute code, and interact with external services, often without systematic inspection of what they do or why. My research builds the methods, benchmarks, and open-source infrastructure that make AI systems auditable: detecting, diagnosing, and controlling departures from expected behavior, from anomalous data and foundation-model failures to consequential agent actions.

Methodologically, this work extends my prior research on anomaly and outlier detection (the basis of the PyOD ecosystem) from data distributions to foundation-model behavior and agent decision traces, where unsafe, anomalous, or out-of-policy actions must be detected and reconstructed before and after deployment.

Three levels in the lab taxonomy:

• 🤖 Agent Action: Auditability and Control of Agent Systems. Auditability frameworks (Auditable Agents, auditable), runtime control that intercepts tool calls before they fire (Aegis), and static over-privilege scanning (agent-audit) with its benchmark (FORTIS). Also post-run failure localization in multi-agent systems (GRADE) and agent-specific failure modes: over-privilege, cross-user contamination, and the autonomy tax of defense training.
• 🧠 Model Behavior: Trust and Robustness of Foundation Models. Jailbreak detection for vision-language models, causal analysis of hallucination, query-agnostic attacks on retrieval-augmented generation, and LLM-as-anomaly-detector benchmarks.
• 📊 Input Distribution: Anomaly and Out-of-Distribution Detection. PyOD ecosystem, ADBench, automatic OOD detector selection, modality-specific OOD methods, and few-shot cross-domain OOD detection.

Our position paper Auditable Agents organizes this into five dimensions of auditability and three mechanism classes (detect, enforce, recover). It appears at the inaugural ACM AI Leadership Summit and at the ACL 2026 KnowFM workshop.

29k+ GitHub stars and 60M+ downloads across the projects I lead or co-author (August 2026). Featured below (full list on the homepage):

Project	Research level / role	Stars	What It Does
PyOD	Input distribution		Canonical anomaly detection, now in its third major release. 60+ detectors across tabular, time series, graph, text, and image data, plus an ADEngine orchestration core and an agentic investigation layer driven in natural language.
auditable	Agent action		System of record for AI-agent decisions: capture, replay against live state, roll back.
GRADE	Agent action		The method auditable is built on: a typed two-layer graph of agent execution and dependency. arXiv
agent-audit	Agent action		Static security and over-privilege scanner for AI-agent code.
awesome-auditable-ai	Agent action		Curated map of papers, tools, datasets, and standards for reliable, auditable AI agents.
anywhere-agents	Supporting tooling		One config for Claude Code and Codex across every project and session.
agent-style	Supporting tooling		21 writing rules loaded into AI agents at generation time.

I lead the FORTIS Lab at USC, working on auditable AI systems, anomaly detection, and foundation-model trust, with a team of PhD, master's, and undergraduate researchers.

My awards include the NVIDIA Academic Grant Program Award, multiple Amazon Research Awards, the Capital One Research Award, and the Foresight Institute AI for Safety & Science Nodes Grant. NSF grants and a Meta sponsored research agreement fund the lab alongside them. AAAI selected me for New Faculty Highlights.

Founder & Maintainer of 博士栈 · CSPhD.org, a non-profit community for CS / AI / EE / Stats PhDs. Since 2022 it has grown to several thousand members and helped many land PhD positions, internships, and full-time roles, with a searchable board of 500+ PhD, RA, and postdoc openings.

• 🌐 Homepage · Google Scholar · LinkedIn
• ✉️ yue.z [AT] usc.edu