Jongseong Kim
First-year M.S. student, Computer Science
University of Illinois Urbana-Champaign
I am a first-year M.S. student in Computer Science at UIUC, supervised by Lingming Zhang. Before graduate school I worked as an offensive security researcher on Windows internals.
My work is on AI agents that find bugs in Windows systems, benchmarks that measure how well they do it, and the tooling that makes both possible. Microsoft has recognized this work with listings on the Most Valuable Researcher leaderboard.
- CVEs assigned
- 133
- MSRC 2026 annual
- #32
- MSRC quarterly MVR
- 7×
- DEF CON CTF Finals
- 3rd
News
- Released DeepBinary, a practical benchmark that measures how well models find bugs in Windows.
- CyberDual is out on arXiv: AI attackers versus AI code reviewers in real software maintenance.
- KEX-bench appears at EMNLP 2026 as a main-conference oral.
- 26 CVEs credited in Microsoft's October Patch Tuesday.
- Started the M.S. in Computer Science at UIUC.
- 1st place at the DEF CON 34 CTF Qualifier, out of 673 teams.
- Agentic Vulnerability Reasoning on COTS Binaries is out on arXiv.
Publications
-
CyberDual: AI Attackers versus AI Code Reviewers in Real Software MaintenancearXiv preprint arXiv:2610.34435, 2026
An arena where attacker agents hide exploitable bugs in real open-source maintenance work and reviewer agents try to catch them, judged by rerunning each proof of concept.
@article{yang2026cyberdual, title = {CyberDual: AI Attackers versus AI Code Reviewers in Real Software Maintenance}, author = {Yang, Chenyuan and Lee, Hwiwon and Kim, Jongseong and Yang, Yihan and Kim, Dongjun and Chen, Yinfang and Zhang, Lingming}, journal = {arXiv preprint arXiv:2610.34435}, year = {2026} } -
Evaluating Coding Agents on Kernel Exploit GenerationProceedings of EMNLP 2026 (Main Conference), Budapest
KEX-bench: 45 tasks from 40 Linux and Windows kernel CVEs that ask coding agents to build exploit primitives, from address leaks to arbitrary writes.
@inproceedings{jang-etal-2026-evaluating, title = {Evaluating Coding Agents on Kernel Exploit Generation}, author = {Jang, Junyoung and Lee, Gwanhyun and Lee, Hwiwon and Kim, Kyuheon and Kim, Jongseong and Jung, Jinho and Zhang, Lingming}, booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing}, pages = {37581--37603}, year = {2026}, address = {Budapest, Hungary}, publisher = {Association for Computational Linguistics}, url = {https://aclanthology.org/2026.emnlp-main.1653/} } -
Agentic Vulnerability Reasoning on COTS BinariesarXiv preprint arXiv:2605.05000, 2026
SLYP, an early prototype of SEC-agent, pairs tool servers for binary exploration and dynamic debugging with debugger-verified proof-of-concept synthesis. On a 20-object COM benchmark it recovers all 64 vulnerable entry functions, while default production coding agents miss up to 15 and produce no verified PoCs.
@article{lee2026agentic, title = {Agentic Vulnerability Reasoning on COTS Binaries}, author = {Lee, Hwiwon and Kim, Jongseong and Zhang, Lingming}, journal = {arXiv preprint arXiv:2605.05000}, year = {2026} }
* Equal contribution
Talks
-
COM-pletely Unplanned: A Windows Bug Hunter's Journey to LPEOff-By-One 2025, Singapore
-
1-Click-Fuzz: Systematically Fuzzing the Windows Kernel Driver with Symbolic ExecutionCODE BLUE 2024, Tokyo
Vulnerability disclosures
My reports have received 133 CVE assignments from Microsoft, Apple, and third-party driver vendors between 2022 and 2026, spanning Windows services and kernel drivers, Office, SharePoint, and Exchange. Microsoft has recognized these reports with seven quarterly Most Valuable Researcher listings and a #32 ranking on the 2026 annual leaderboard.
Browse all 133 CVEs