About

I am a Research Scientist in the AI Supply Chain Security Group at IBM Research. My research areas are software and systems security, program analysis, AI for security, autonomous vulnerability discovery, secure software supply chains, and the security of AI agents.

My work began in OS kernels and open-source infrastructure, and it now also covers AI-generated code and the agents that write and analyze it. The Research section below describes those lines and the systems that came out of them.

I received my Ph.D. in Computer Science and Engineering from the University of Minnesota, Twin Cities (advisor: Kangjie Lu), and my B.Eng. in Information Security from the University of Science and Technology of China.

Research

Agents now write and modify a growing share of the code everyone depends on, faster than it can be reviewed. Most vulnerability detection still needs someone to write down what the bug looks like first, which limits it to the classes people have already thought of.

I build methods that recover the expectation from the software itself: from how the code behaves, from what comparable code does, from what a fix changed. Kernels, cryptographic libraries, compilers, and now AI-written code have been the targets, but nothing in the methods is specific to them. A method that reports more also has to say which of its reports are real, so much of the work is confirmation: executing the path, and sampling and auditing findings by hand.

Static analysis for memory and logic bugs in systems software

The rules are specific to the codebase: what has to be checked before a pointer is used, what has to be released on an error path, what one function promises the next. I build analyses that infer those rules from the code itself and flag every place they are broken. I built HERO (USENIX Security 2021), which pairs functions to detect disordered error handling, and GNNIC (NDSS 2024), which recovers indirect-call targets and is the basis of a granted U.S. patent. With collaborators I worked on CRIX, IPPO, K-MeLD, and NDI, which catch missing checks, missed cleanup, memory leaks, and inconsistent paths.

Fuzzing that reaches hard-to-trigger code

A static warning stays a lead until something executes the path. With collaborators I worked on driver fuzzing that infers device semantics and needs neither the hardware nor an emulator. I also contributed to directed fuzzing guided by feasibility, and to work on exposing which crashes are exploitable.

Learning from security patches to prioritize and classify vulnerabilities

Most security fixes land quietly, so what a bug allowed is buried in a diff and a terse commit message. I built SID (NDSS 2020), which compares symbolic rules before and after a patch to decide whether it fixes a security bug and what that bug allowed. I co-led DiffCVSS (USENIX Security 2022), which shows that one vulnerability deserves different severity scores on different operating systems. I contributed to DualLM (NDSS 2026), which categorizes kernel patches by the memory bugs they fix.

LLM agents for validated bug discovery, and the security of AI-written code

A fixed bug describes a mistake that other code is still making. I led BugStone (ICML 2026), which turns fixed bugs into reusable detection rules. In the Linux kernel, 135 rules flagged 22,568 candidate sites, and 246 of 400 sampled at random were confirmed by hand; on a separate 850-patch benchmark, the rules reached 92.2% precision. Model-written code needs the same scrutiny, and I contributed to APILOT (ACSAC 2025), which steers LLM code suggestions away from outdated and vulnerable APIs. I now build agents that find, validate, and report vulnerabilities with evidence a maintainer can check, and I work on red-teaming agent systems themselves. ClawStack is a public experiment in the same direction: agents with separated roles, policy enforcement, and audited execution. Throughout, I would rather a static analysis answer a question than a model, because the analysis is cheaper and its answer can be checked.

News

2026
“One Bug, Hundreds Behind: LLMs for Large-Scale Bug Discovery” appears at ICML 2026.
2026
“What Do They Fix? LLM-Aided Categorization of Security Patches for Critical Memory Bugs” appears at NDSS 2026.
2025
APILOT is accepted at ACSAC 2025 and awarded ACSAC’s Code Reviewed badge.
2025
Serving on the NDSS 2025 program committee.
2024
Presented GNNIC at NDSS 2024.
2024
Co-organized the AISCC workshop, co-located with NDSS 2024, and served on the IEEE S&P 2024 program committee.
Dec 2023
U.S. Patent 11,853,751, “Indirect Function Call Target Identification in Software,” is granted.
Jul 2023
Joined IBM Research as a Research Scientist.

Publications

Peer-reviewed conference and journal papers, with workshop papers and preprints marked inline. * denotes co-first authors.

2026

  1. One Bug, Hundreds Behind: LLMs for Large-Scale Bug Discovery
    Qiushi Wu, Yue Xiao, Dhilung Kirat, Kevin Eykholt, Jiyong Jang, and Douglas Lee Schales
    ICML 2026 · International Conference on Machine Learning [paper] [arXiv] [dataset]
  2. What Do They Fix? LLM-Aided Categorization of Security Patches for Critical Memory Bugs
    Xingyu Li, Juefei Pu, Yifan Wu, Xiaochen Zou, Shitong Zhu, Qiushi Wu, Zheng Zhang, Joshua Hsu, Yue Dong, Zhiyun Qian, Kangjie Lu, Trent Jaeger, Michael De Lucia, and Srikanth V. Krishnamurthy
    NDSS 2026 · Network and Distributed System Security Symposium [paper] [code]

2025

  1. APILOT: Improving the Security and Usability of LLM Code Suggestions via Outdated API Mitigation
    Weiheng Bai, Keyang Xuan, Pengxiang Huang, Qiushi Wu, Jianing Wen, Jingjing Wu, and Kangjie Lu
    ACSAC 2025 · Annual Computer Security Applications Conference · Code Reviewed badge [paper] [code]
  2. AFLGopher: Accelerating Directed Fuzzing via Feasibility-Aware Guidance
    Weiheng Bai, Kefu Wu, Qiushi Wu, and Kangjie Lu
    arXiv 2025 · Preprint [arXiv]

2024

  1. GNNIC: Finding Long-Lost Sibling Functions with Abstract Similarity
    Qiushi Wu, Zhongshu Gu, Hani Jamjoom, and Kangjie Lu
    NDSS 2024 · Network and Distributed System Security Symposium [paper]
  2. Exploring the Influence of Prompts in LLMs for Security-Related Tasks
    Weiheng Bai, Kefu Wu, Qiushi Wu, and Kangjie Lu
    NDSS Workshop 2024 · AISCC, co-located with NDSS · Workshop paper [paper]

2023

  1. Silent Bugs Matter: A Study of Compiler-Introduced Security Bugs
    Jianhao Xu, Kangjie Lu, Zhengjie Du, Zhu Ding, Linke Li, Qiushi Wu, Mathias Payer, and Bing Mao
    USENIX Security 2023 · USENIX Security Symposium [paper]
  2. Guiding Directed Fuzzing with Feasibility
    Weiheng Bai, Kefu Wu, Qiushi Wu, and Kangjie Lu
    EuroS&P Workshops 2023 · IEEE European Symposium on Security and Privacy Workshops · Workshop paper [paper]
  3. Towards More Effective Responsible Disclosure for Vulnerability Research
    Weiheng Bai and Qiushi Wu
    EthiCS 2023 · Workshop on Ethics in Computer Security · Workshop paper [paper]

2022

  1. Non-Distinguishable Inconsistencies as a Deterministic Oracle for Detecting Security Bugs
    Qingyang Zhou, Qiushi Wu, Dinghao Liu, Shouling Ji, and Kangjie Lu
    CCS 2022 · ACM Conference on Computer and Communications Security [paper] [code]
  2. OS-Aware Vulnerability Prioritization via Differential Severity Analysis
    Qiushi Wu*, Yue Xiao*, Xiaojing Liao, and Kangjie Lu (*co-first authors)
    USENIX Security 2022 · USENIX Security Symposium [paper] [code & data]
  3. Semantic-Informed Driver Fuzzing Without Both the Hardware Devices and the Emulators
    Wenjia Zhao, Kangjie Lu, Qiushi Wu, and Yong Qi
    NDSS 2022 · Network and Distributed System Security Symposium [paper]

2021

  1. Detecting Missed Security Operations Through Differential Checking of Object-based Similar Paths
    Dinghao Liu, Qiushi Wu, Shouling Ji, Kangjie Lu, Zhenguang Liu, Jianhai Chen, and Qinming He
    CCS 2021 · ACM Conference on Computer and Communications Security [paper]
  2. Detecting Disordered Error Handling with Precise Function Pairing
    Qiushi Wu, Aditya Pakki, Navid Emamdoost, Stephen McCamant, and Kangjie Lu
    USENIX Security 2021 · USENIX Security Symposium [paper]
  3. Practically Detecting Kernel Memory Leaks in Specialized Modules and Beyond
    Navid Emamdoost, Qiushi Wu, Kangjie Lu, and Stephen McCamant
    NDSS 2021 · Network and Distributed System Security Symposium [paper]
  4. Unleashing Fuzzing Through Comprehensive, Efficient, and Faithful Exploitable-Bug Exposing
    Bowen Wang*, Kangjie Lu*, Qiushi Wu, and Aditya Pakki (*co-first authors)
    IEEE TDSC 2021 · IEEE Transactions on Dependable and Secure Computing · Journal [paper]

2020

  1. Precisely Characterizing Security Impact in a Flood of Patches via Symbolic Rule Comparison
    Qiushi Wu, Yang He, Stephen McCamant, and Kangjie Lu
    NDSS 2020 · Network and Distributed System Security Symposium [paper]

2019

  1. Detecting Missing-Check Bugs via Semantic- and Context-Aware Criticalness and Constraints Inferences
    Kangjie Lu, Aditya Pakki, and Qiushi Wu
    USENIX Security 2019 · USENIX Security Symposium [paper] [code]
  2. Automatically Identifying Security Checks for Detecting Kernel Semantic Bugs
    Kangjie Lu, Aditya Pakki, and Qiushi Wu
    ESORICS 2019 · European Symposium on Research in Computer Security [paper]

Software & Artifacts

  • BugStoneBench (first-authored, ICML 2026) — benchmark dataset accompanying “One Bug, Hundreds Behind,” for evaluating LLM-based large-scale bug discovery. [dataset]
  • DIFFCVSS (co-first-authored, USENIX Security 2022) — prioritizes vulnerabilities via OS-aware differential severity analysis. [code & data]
  • APILOT (co-authored, ACSAC 2025) — mitigates outdated, vulnerable API usage in LLM code suggestions; awarded ACSAC’s Code Reviewed badge. [code]
  • DualLM (co-authored, NDSS 2026) — detects security-critical Linux kernel patches, focusing on use-after-free and out-of-bounds vulnerabilities. [code]
  • CRIX (co-authored, USENIX Security 2019) — detects missing-check bugs in OS kernels. [code]
  • NDI (co-authored, CCS 2022) — a deterministic oracle for detecting security bugs. [code]
  • ClawStack — public experiments applying AI agents to real CTF-style security tasks; top 6% at BearcatCTF 2026 (40 of 44 challenges solved). [showcase]

Patents

  • Qiushi Wu, Zhongshu Gu, and Hani Jamjoom. “Indirect Function Call Target Identification in Software.” U.S. Patent 11,853,751.
  • Qiushi Wu, Zhongshu Gu, Enriquillo Valdez, and Hani Jamjoom. “Detecting Bugs Using Large Language Models.” U.S. Patent Application 2026/0017174 A1 (pending).

Research Ethics

In 2021, as a Ph.D. student at the University of Minnesota, I was the first author of the hypocrite commits work. We sent Linux kernel maintainers patches designed to introduce security bugs, from pseudonymous accounts and without their knowledge or consent, to test whether review would catch them. That cost maintainers time they had not agreed to give. The design was wrong. I am sorry. The primary record is public: my department’s collected statements and correspondence, the Linux Foundation’s Technical Advisory Board report, our apology, and the withdrawal letter.

What came after is the part I can speak for. The group I was in took research ethics up as a line of work, and my advisor documents what came of it. My own part was a 2023 workshop paper on responsible disclosure, written with a student I mentored. A Linux kernel patch I sent in 2025 was reviewed, revised at the maintainer’s request, and merged. For vulnerability research affecting external software I validate the finding, notify the affected maintainers before public release, and coordinate disclosure. For research involving developer behavior I require explicit opt-in consent and full disclosure, consistent with the kernel’s researcher guidelines. I welcome questions about the 2021 study or my current practice.

Teaching & Mentoring

  • CSCI 4271: Development of Secure Software Systems (University of Minnesota, Fall 2022) — Graduate Teaching Assistant; sole graduate TA for a class of approximately 40–60 students; designed hands-on labs, including a vulnerability-finding lab and an exploitation lab.
  • CSCI 2021: Machine Architecture and Organization (University of Minnesota, Fall 2018) — Graduate Teaching Assistant; one of two graduate TAs for 200+ enrolled students, coordinating a team of 5+ undergraduate TAs.
  • Mentoring — mentored a junior Ph.D. student through multiple co-authored publications (now an Applied Scientist at Amazon); currently mentoring an IBM Research Ph.D. intern on the security of agentic systems and co-mentoring two first-year Ph.D. students on LLM/agentic methods for security.

Service

  • Program committee: IEEE Symposium on Security and Privacy (IEEE S&P), 2024; Network and Distributed System Security Symposium (NDSS), 2025.
  • Organizing committee: Workshop on Artificial Intelligence System with Confidential Computing (AISCC), co-located with NDSS, 2024.
  • External reviewer: ACM CCS 2019/2020; ICICS 2019; NDSS 2021.