Dataisland / security research
Zheng Yu 于峥
I build agents and systems for finding, understanding, and repairing vulnerabilities.
Security Researcher at DepthFirst, working on LLM-based agents for vulnerability discovery, repair, and penetration testing.
PhD in Computer Science from Northwestern University; bachelor's degree from the ACM Honor Class at Shanghai Jiao Tong University.
-
pPatch accepted to ACM FSE 2026.
-
Joined DepthFirst as a Security Research Engineer.
-
Passed PhD defense at Northwestern.
-
I got 🥈 Runner-up in Technical Impact Award of CSAW Applied Research Competition (ARC) 2025!
-
PC member for LLM4Code 2026 at ICSE.
-
Passed PhD proposal exam.
-
PatchAgent selected as a CSAW ARC 2025 finalist.
-
PortGPT accepted to IEEE S&P 2026.
-
Our team, 42-b3yond-6ug, won rank 6 at AIxCC Finals!
-
Awarded Student Grant from USENIX Security 2025!
-
PatchAgent accepted to USENIX Security 2025.
-
Received MS in Computer Science from Northwestern.
-
Passed PhD qualifying exam.
-
Awarded Student Grant from ACM CCS 2024!
-
Our team, 42-b3yond-6ug, won the AIxCC Semi-Finals and was awarded $2 million. I had the honor of designing the program repair system for the competition.
-
Awarded Student Grant from USENIX Security 2024!
-
LLM-Fuzzer accepted to USENIX Security 2024.
-
ShadowBound accepted to USENIX Security 2024.
-
Our team, 42-b3yond-6ug, received funding of $1 million from AIxCC.
-
CAMP accepted to USENIX Security 2024.
-
DepthFirst / Security Researcher AI Security Team
-
CertiK / Security Researcher Smart Contract Security Team
-
University of Waterloo / Visiting Researcher CrySP Lab, Advisor: Prof. Meng Xu
-
Google Summer of Code / GSOC Project Mentor Bridging Qiling and Static Analysis
-
JD.COM, Inc. / Security Engineer Application Security Team - Firmware Emulation
-
Shanghai Jiao Tong University / Infrastructure Engineer Information and Network Center
-
Right Code, Wrong Comment: The Pointer Chase Inside zlib
A 38-line comment in deflate.c argues that zlib's shared buffer can never overwrite unread symbols. That comment is wrong—yet when we tried to build an exploit, three hidden Huffman invariants stood in the way.
-
Inside the V8 Sandbox
Notes on V8's sandbox threat model, attack boundary, and the browser code that still has to treat sandbox data carefully.
-
Bitcoin From the Top Down
A top-down walkthrough of Bitcoin's data model, from peer-to-peer state to transactions, UTXOs, and script validation.
- pPatch: Automated Vulnerability Unpatching ACM International Conference on the Foundations of Software Engineering (FSE) 2026
- PortGPT: Towards Automated Backporting Using Large Language Models IEEE Symposium on Security and Privacy (S&P) 2026
- PatchAgent: A Practical Program Repair Agent Mimicking Human Expertise USENIX Security Symposium 2025 (Long Presentation)
- ShadowBound: Efficient Heap Memory Protection Through Advanced Metadata Management and Customized Compiler Optimization USENIX Security Symposium 2024
- LLM-Fuzzer: Scaling Assessment of Large Language Model Jailbreaks USENIX Security Symposium 2024
- Reversing MCU with Firmware Emulation BlackHat Europe 2022, London, UK