Preprint · arXiv 2026
Chenyuan Yang*, Hwiwon Lee*, Jongseong Kim, Yihan Yang, DongJun Kim, Yinfang Chen, Lingming Zhang
Agents act as attackers that embed exploitable vulnerabilities in open-source code changes and as reviewers that try to catch and repair them; re-running the attacker's proof of concept decides whether the exploit survives. 1,088 valid attacks from six attackers against three agentic reviewers.
Benchmark · 2026
Jongseong Kim*, Hwiwon Lee*, DongJun Kim, Lingming Zhang
Benchmarks AI agents on real Windows system binaries, scoring findings by executable evidence, security impact, and estimated bounty. Every zero-day found through this work has been reported to MSRC.
Preprint · arXiv 2026
Hwiwon Lee, Jiawei Liu, DongJun Kim, Ziqi Zhang, Chunqiu Steven Xia, Lingming Zhang
Long-horizon security tasks in million-line systems — V8, SpiderMonkey, and the Linux kernel — including vulnerabilities with over $1 million in cumulative bounties. Adopted for frontier-model evaluation by OpenAI (GPT-5.5-Cyber, GPT-6 Astra), Google (Gemini 4 Carbon), DeepSeek, and NIST CAISI.