Agentic security
Autonomous agents for realistic, long-horizon security engineering.
Slide 01 · Software Security × AI
Hwiwon
Lee
PhD student in Computer Science at UIUC · advised by Prof. Lingming Zhang
I build and evaluate AI agents for cybersecurity.
Slide 02 · Inquiry
Research,in motion.
I develop rigorous systems and evaluations that move security agents beyond short demonstrations and into the long, uncertain workflows of real vulnerability research.
Software systems
Model reasoning grounded in program analysis and execution evidence.
Vulnerability research
Discovery and validation across production software and binaries.
Slide 03 · Signals
Recent
news
-
SEC-bench Pro, our benchmark for long-horizon software security tasks, is now available on arXiv.
-
Agentic Vulnerability Reasoning on COTS Binaries is now available on arXiv.
-
Recognized as a Microsoft Most Valuable Security Researcher for 2026 Q1.
Slide 04 · Record
Selected
work
Benchmarks and systems for measuring, grounding, and advancing AI-driven security research.
-
arXiv
2026PDF of SEC-bench ProSEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?
A benchmark of 344 validated vulnerabilities across browser engines and the Linux kernel, designed to measure realistic agent bug hunting.
-
arXiv
2026PDF of Agentic Vulnerability Reasoning on COTS BinariesAgentic Vulnerability Reasoning on COTS Binaries
An end-to-end study of autonomous vulnerability discovery and debugger-verified validation on commercial Windows binaries.
-
NeurIPS
2025PDF of SEC-benchSEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks
A fully automated framework for constructing and evaluating authentic proof-of-concept generation and vulnerability patching tasks.
-
IEEE S&P
2024PDF of BENZENEBENZENE: A Practical Root Cause Analysis System with an Under-Constrained State Mutation
A practical, automated crash-diagnosis system that uses under-constrained state mutation to rank root causes efficiently.