01 / AGENTSACTIVE
Agentic security
Long-horizon agents that investigate complex targets, form hypotheses, and produce verifiable security artifacts.
cat research_statement.md
I build and evaluate AI agents for cybersecurity.
Hwiwon Lee is a Computer Science PhD student at UIUC, advised by Lingming Zhang. His work develops realistic agents, benchmarks, and systems for long-horizon software security tasks.
cat research_areas.yml
Research modules
I study security agents as complete systems: how they reason, how they operate on real software, and how we can evaluate their progress with evidence rather than toy success metrics.
02 / SYSTEMSACTIVE
Software systems
Infrastructure for reproducible experiments, realistic tool use, and dependable security engineering workflows.
03 / EVIDENCEACTIVE
Vulnerability research
Methods grounded in real binaries and failure modes, with outcomes that matter to researchers and practitioners.
tail -n 3 news.log
Recent signal
New releases and recognition from the research log.
SEC-bench Pro released on arXiv.
Agentic Vulnerability Reasoning on COTS Binaries released on arXiv.
Recognized as a Microsoft Most Valuable Security Researcher, 2026 Q1.
ls papers/ --selected
Selected work
Research spanning autonomous vulnerability reasoning, realistic agent evaluation, and software root-cause analysis.
01
SEC-bench Pro: Can Language Models Solve Long-Horizon Software Security Tasks?
arXiv
2026 02 Agentic Vulnerability Reasoning on COTS Binaries arXiv
2026 03 SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks NeurIPS
2025 04 BENZENE: A Practical Root Cause Analysis System with an Under-Constrained State Mutation IEEE S&P
2024
2026 02 Agentic Vulnerability Reasoning on COTS Binaries arXiv
2026 03 SEC-bench: Automated Benchmarking of LLM Agents on Real-World Software Security Tasks NeurIPS
2025 04 BENZENE: A Practical Root Cause Analysis System with an Under-Constrained State Mutation IEEE S&P
2024