Xiangwei Zhang
中文
Photo of Xiangwei Zhang

Xiangwei Zhang张向伟

I joined Tencent Security Yunding Lab in 2024 as a senior security researcher. Most of my work is finding vulnerabilities: I have found and reported bugs in browser engines, network protocol parsers, and AI infrastructure, and have been acknowledged by Apple, Google, Microsoft, Amazon, Meta, NVIDIA, Mozilla, and others. I work with the offense and defense team on exploit development, and on attack-surface assessments of internal products such as Tencent Meeting and QQ Browser. I also worked on the eBPF host-monitoring plugin for Onion, Tencent’s host intrusion detection system, covering on the order of a million hosts. I now work on capability development for CodeBuddy Security, Tencent’s AI code security audit platform. I studied at Tianjin University from 2017 to 2024 (B.S. and M.S.), advised by Prof. Junjie Wang, fuzzing JavaScript engines and WebAssembly compilers.

Background

Experience and education

  1. 2024.04 — present

    Senior Security Researcher

    Tencent Security Yunding Lab

    AI code auditing: CodeBuddy Security

    2025.12 — present
    • Designed and built VulnAgent, a vulnerability-discovery agent: intelligence scopes the search, the agent audits module by module, an independent pass reviews, and a PoC in a sandbox confirms. The first version ran end to end within two weeks.
    • Turned VulnAgent into the AI deep-audit engine of the CodeBuddy Security platform: batch orchestration and concurrency, multi-round audits that converge without repeating, external intelligence input, incremental diff review, and automated PoC verification.
    • 16 public CVEs found with it, in React, NVIDIA Megatron-LM, Apache IoTDB, Suricata, FreeRDP, LiteLLM, and others. On a benchmark that scores recall of historical CVEs in well-known open-source projects, it reached 73.2% end to end.

    Vulnerability research and red-team support

    2024.08 — 2025.12
    • Built a generation-based WebAssembly fuzzer and ran it against JavaScriptCore, V8, and SpiderMonkey: CVE-2024-54508 (WebKit), CVE-2025-1933 (Firefox), and further JSC and V8 bugs reported to the vendors.
    • Worked with the offense and defense team: turned public browser n-days into reliable exploits across dozens of versions, and found and weaponized vulnerabilities in widely used desktop clients for red-team exercises.
    • Attack-surface assessments of internal products such as Tencent Meeting and QQ Browser, and of LLM applications, AI code sandboxes, and MCP tool chains during offense and defense exercises, with hardening guidance delivered to the product teams.
    • Adapted and shipped C/C++ rules for the in-house static analysis platform and worked down false positives during rollout. Patch analysis and fuzzing of network protocol parsers in this period became the starting point for VulnAgent.

    Host intrusion detection and cloud security

    2024.04 — 2025.04
    • Worked on development of the eBPF host-monitoring plugin for Onion, Tencent’s host intrusion detection system. I took over testing and tuning of the DNS query monitor and added container attribution, built the TCP scan-behavior monitor that ties scans to the originating process, and wrote YARA rules for common proxy tools and their variants.
    • Ran the plugin’s staged rollout and release monitoring; it now runs on the order of a million hosts.
    • Security testing of cloud host and network agents: found command injection, SQL injection, and arbitrary file upload issues, and helped the owning teams fix them.

  2. 2017 — 2024

    B.S. and M.S.

    Tianjin University

    Software security, advised by Prof. Junjie Wang. Fuzzing for JavaScript engines and WebAssembly compilers: PatchFuzz and WasmCFuzz, and a contribution to FuzzJIT.

Featured project

CodeBuddy Security

Details →

CodeBuddy Security is Tencent’s AI code security audit platform, jointly developed by Yunding Lab, Keen Lab, and the Tencent Cloud Code Analysis (TCA) team. I work on the platform’s capability side. The details page covers how the engine finds vulnerabilities and how it was evaluated.

Internal preview · as of 12 Oct 2026
8M
model calls
25,000+
vulnerability findings
3,200+
vulnerabilities reproduced in real environments with engine-written PoCs

Vulnerabilities

Selected CVEs

All CVEs (30)

Major vendors involved, all confirmed and fixed: Apple · Meta · NVIDIA · Mozilla · Apache · OISF.

Research

Publications

  1. PatchFuzz: Patch Fuzzing for JavaScript Engines

    Junjie Wang, Zhihua Xie, Xiaofei Xie, Xiaoning Du, Xiangwei Zhang

    Information and Software Technology, vol. 194, 2026

    Journal article 108087. Preprint: arXiv:2505.00289.

    PaperarXivCode

    Workflow diagramWorkflow of PatchFuzz
  2. WasmCFuzz: Structure-aware Fuzzing for Wasm Compilers

    Xiangwei Zhang, Junjie Wang, Xiaoning Du, Shuang Liu

    EnCyCriS/SVM '24, pp. 1–5

    Proceedings of the 2024 ACM/IEEE 4th International Workshop on Engineering and Cybersecurity of Critical Systems (EnCyCriS) and 2024 IEEE/ACM Second International Workshop on Software Vulnerability

    Paper

    Workflow diagramWorkflow of WasmCFuzz
  3. An Empirical Study on AST-level Mutation-based Fuzzing Techniques for JavaScript Engines

    Song Tang, Shuang Liu, Junjie Wang, Xiangwei Zhang

    Internetware 2023

    Paper

Earlier research

  1. FuzzJIT: Oracle-Enhanced Fuzzing for JavaScript Engine JIT Compiler

    Junjie Wang, Zhiyi Zhang, Shuang Liu, Xiaoning Du, Junjie Chen

    USENIX Security 2023, pp. 1865–1882

    Authors are my advisor Junjie Wang and collaborators. I was responsible for the experimental part, and used the tool to find several bugs.

    PaperCode

    Workflow diagramWorkflow of FuzzJIT