MemHarmBench was recommended by Shanghai Innovation Institute and featured in MiraclePlus's Alpha Sight.
About Me
Hello! I am Zhihua Wang, an incoming CS master's student at Fudan University's Vision and Learning Laboratory (FVL), led by Prof. Yu-Gang Jiang (Vice President, Fudan University), and the Institute of Trustworthy Embodied Artificial Intelligence (TEAI). I will be advised by Prof. Xingjun Ma. I am pursuing a bachelor's degree in Computer Science and Technology at the School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology (USST) (2023–2027). My open-source projects have received approximately 2.5k GitHub Stars, and I won an ACM silver medal.
I am currently an LLM research intern at the Third Research Institute of the Ministry of Public Security (TRIMPS), guided by Nan Li, working on Web Agent and LLM safety evaluation benchmarks. I also collaborate with Yutao Wu, a PhD student at Deakin University, and Yunhao Feng, a PhD student interning at (Ant Group), on multi-agent systems, coding agents, and trustworthy AI.
My research interests include foundation models, coding agents, agent post-training and reinforcement learning, and trustworthy AI. 🙂(:
News
The code for MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI Agents is publicly available on GitHub, and the paper is available on arXiv.
The code for CST: Lightweight Risk Triage with Source-Grounded Evidence for Web Discussions is publicly available on GitHub, and the paper is available on arXiv.
Published a Logitech mouse sleep/wake investigation and recovery scripts, documenting receiver power-loss behavior and the initial filter-driver hypothesis that later testing ruled out. Updated the attribution in the upstream issue.
Participated in the open-source projects AISafetyHot and prompt-spider-ui during China's National Day holiday. Many thanks to my collaborators for their support and ideas.
Received the National First Prize at the National Finals of the 9th National College Student Embedded Chip and System Design Competition.
Completed research on multimodal data curation, grounded annotation, benchmark design, and training-free fraud understanding for ScamWeb: A Multimodal Benchmark for Grounded Understanding of Cyber-enabled Fraud Webpages.
Joined Advisor Jingfeng Zhang's RISE-AI Lab as a visiting student at TEAI, Fudan University and began research on MemHarmBench, a benchmark for persistent-memory poisoning in agents. The work adopts a victim-centered perspective to design evaluation cases and examines who ultimately bears the loss after an agent is compromised.
Began serving as a reviewer for NeurIPS 2026, contributing to the peer-review process for research in machine learning and artificial intelligence.
At TEAI, Fudan University, under the guidance of Advisor Xingjun Ma, I participated in an enterprise-oriented multi-agent secure collaboration project, contributing to requirements analysis, system design, implementation, and testing.
Completed an International Organization Internship through the UNDP-administered UNV platform on mangrove responses to climate change and produced a review manuscript on blue carbon mitigation.
Joined TEAI, Fudan University and, under the guidance of Advisor Xingjun Ma, worked on the XSafeClaw agent safety project, focusing on runtime integration, controlled tool calls, and safety audit workflows.
Started SafeCodeRL, a multi-agent reinforcement learning framework for safety-constrained LLM code generation.
Worked on VulnSeeker at the Fudan Software Engineering (CodeWisdom) Lab, combining CodeQL static analysis with LLM-based security judgement.
Joined the Fujian Key Laboratory of Network Computing and Intelligent Information Processing at Fuzhou University and, under the guidance of Advisor Yizhang Liu, began SFMambaNet research on spectral-frequency enhanced state space models for correspondence pruning.
My chapter as an ACMer officially came to an end.
Joined the ACM Programming Training Team at the University of Shanghai for Science and Technology and began my ACM journey. Looking forward to the challenges ahead! 🚀
AISafetyHot
An open-source project I participated in during the National Day holiday. It brings together AI safety news, research, and events, with an MCP interface for agents.
Demo credit: AI Safety HOT · README Explore the website ↗
Preprints and Publications
AAAI 2027 · CCF A ScamWeb: A Multimodal Benchmark for Grounded Understanding of Cyber-enabled Fraud Webpages
Co-first author · AAAI 2027; Will be accepted(accept*4)
Preprint MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI Agents
First author · arXiv preprint
Preprint CST: Lightweight Risk Triage with Source-Grounded Evidence for Web Discussions
First author · arXiv preprint
SCI Zone 3 · CCF C SafeCodeRL: Security-Constrained Multi-Agent Reinforcement Learning for Trustworthy LLM-Generated IoT/CPS Software
First author · Sensors, 26(11), 3502
SCI Zone 1 · CCF A Smooth Inliers, Erratic Outliers: Frequency-Aware Mamba for Two-View Correspondence Learning
First author · Under review at IEEE Transactions on Image Processing (TIP) · arXiv preprint
Preprint Research Support in Analysing Mangrove Responses to Climate Change: The Interplay of Anthropogenic Impacts and Blue Carbon Mitigation
Sole author · UNDP / UNV review manuscript
Educations
- 2027.09 – 2030.06 (expected) Fudan University · Incoming CS M.S. Student
- 2023.09 – 2027.06 University of Shanghai for Science and Technology · B.Eng. in Computer Science and Technology (in progress)
Work Experience
AI Research Intern, Key Laboratory of Information Network Security, Ministry of Public Security (Third Research Institute, Shanghai)
- Advisor: Songyang Wu
- Project 1: Independently built a pipeline for joint analysis of webpage text, images, and URLs and localization of fraud evidence, producing one sole-first-author paper.
- Project 2: Designed a lightweight algorithm for short-text risk-topic discovery, combining character and frozen semantic features to aggregate risk evidence from group chats and enable rapid CPU-based screening.
International Organization (United Nations Volunteers) - UNDP Data Research Intern (Online)
Meituan, Odd Jobs
Selected Research and Experience
2026.08 – 2026.11 Built MemHarmBench for cross-session memory poisoning and CST for CPU-based risk triage.
Research Intern, MemHarmBench and CST
Institute of Trustworthy Embodied AI (TEAI), Fudan University · FVL Laboratory · Shanghai Innovation Institute
- Advised by Xingjun Ma and Yu-Gang Jiang (Vice President, Fudan University).
- MemHarmBench: Built an exposure–trigger benchmark for cross-session memory poisoning with 288 case environments (24 attack templates × 12 business scenarios) and 12 target categories. On samples retaining malicious rules, indirect attack success reached 59.38%–76.39%.
- CST: Designed a lightweight algorithm for short-text risk-topic discovery, combining character-level and frozen semantic features to aggregate risk evidence from group chats for rapid CPU screening and source-grounded review.
2026.04 – 2026.08 Co-developed a multilingual fraud benchmark with expert-annotated evidence and explanations.
Research Intern, ScamWeb Multimodal Fraud Benchmark
Institute of Trustworthy Embodied Artificial Intelligence (TEAI), Fudan University & The Third Research Institute of the Ministry of Public Security
- Conducted joint research at the Institute of Trustworthy Embodied Artificial Intelligence (TEAI), Fudan University and The Third Research Institute of the Ministry of Public Security.
- Worked under the guidance of Advisor Xingjun Ma and collaborated with postdoctoral researcher Dr. Nan Li on the full research lifecycle.
- Co-developed ScamWeb, a multimodal and multilingual benchmark containing more than 10,000 curated webpages across 25 cyber-enabled fraud subtypes, including 3,416 expert-annotated fraudulent webpages with evidence regions and human-written rationales.
- Contributed to data collection and cleaning, fraud taxonomy and expert annotation protocol design, the ScamDet multimodal pipeline and prompts, evaluation metrics, full-scale experiments, manuscript writing, result analysis, and project implementation.
- The research manuscript is titled ScamWeb: A Multimodal Benchmark for Grounded Understanding of Cyber-enabled Fraud Webpages.
2026.03 – 2026.06 Integrated local agent sessions with controlled tool calls and runtime safety audits.
Research Intern, XSafeClaw Agent Runtime Safety
Institute of Trustworthy Embodied Artificial Intelligence (TEAI), Fudan University
- Worked on XSafeClaw, an open-source safety platform for local and multi-runtime agent systems, at TEAI, Fudan University.
- Worked under the guidance of Advisor Xingjun Ma, a Young Researcher and doctoral supervisor at Fudan University.
- Implemented runtime integration for nanobot-style local agent sessions, including session discovery, controlled tool calls, and unified platform display.
- Contributed to risk control workflows, validation, and documentation for explainable agent governance.
2026.01 – 2026.03 Combined CodeQL analysis, code context, and LLM judgement for vulnerability triage.
Research Intern, VulnSeeker
Fudan Software Engineering (CodeWisdom) Lab
- Conducted research at the Fudan Software Engineering (CodeWisdom) Lab, focusing on automated security analysis for generated and repository code.
- Worked under the guidance of Lecturer Yi Li and collaborated with Qicai Chen, a Ph.D. student, on literature review, workflow design, and experimental validation.
- Built a CodeQL-assisted vulnerability analysis workflow that combines static analysis, code context expansion, and LLM-based security judgement.
- Designed prompt workflows for CodeQL and LLM interaction, including structured status codes for vulnerability triage and post-generation security inspection.
2026.01 – 2026.04 Reviewed mangrove responses to climate change and blue carbon mitigation through UNV.
International Organization Internship, UNDP-administered UNV Online Research Intern
Morobe Development Foundation Inc., Papua New Guinea
- Completed an International Organization Internship through the UNDP-administered United Nations Volunteers (UNV) platform, supporting research on mangrove responses to climate change scenarios and blue carbon mitigation.
- The United Nations Volunteers (UNV) programme is administered by the United Nations Development Programme (UNDP).
- Synthesized literature on sea-level rise, coastal squeeze, nutrient loading, carbon sequestration, and national climate policy.
- Produced a review manuscript titled Research Support in Analysing Mangrove Responses to Climate Change: The Interplay of Anthropogenic Impacts and Blue Carbon Mitigation.
2026.02 – 2026.05 Designed a five-agent, constraint-aware RL workflow for secure LLM code generation.
First-author Research, SafeCodeRL
Independent research collaboration
- Proposed SafeCodeRL, a multi-agent framework for dynamic safety constraints in LLM code generation.
- Designed a closed-loop collaboration workflow across five agents and a PPO-style constraint-aware policy.
- Reported a large reduction in high-risk vulnerable code generation while preserving functional correctness in the manuscript.
2025.09 – 2026.03 Explored spectral-frequency state space models for robust two-view correspondence pruning.
Research Intern, SFMambaNet
Fujian Key Laboratory of Network Computing and Intelligent Information Processing, Fuzhou University
- Worked as a research intern at Fujian Key Laboratory of Network Computing and Intelligent Information Processing, Fuzhou University (2025.09 – 2026.03).
- Worked closely with Advisor Yizhang Liu, a university-appointed Associate Research Fellow and graduate supervisor, on robust two-view correspondence pruning.
- This work led to our manuscript Smooth Inliers, Erratic Outliers: Frequency-Aware Mamba for Two-View Correspondence Learning, currently under review at IEEE Transactions on Image Processing (TIP) and available as an arXiv preprint.
2024.05 – 2025.07 Trained in algorithms and data structures; earned national and regional programming awards.
ACMer, ACM Programming Training Team
University of Shanghai for Science and Technology
- Completed systematic training in algorithms, data structures, and contest problem solving.
- Earned a National First Prize in the 2025 RAICOM Robot Developer Competition, a Silver Award at the 2025 ACM Shanghai Collegiate Programming Contest, and a National Second Prize in the 2025 LanQiao Cup C/C++ competition.
Awards
- Silver Award, ACM Shanghai Collegiate Programming Contest
- National First Prize, RAICOM Robot Developer Competition,
- National Second Prize, LanQiao Cup C/C++ Programming Contest,
- National First Prize, National Finals of the 9th National College Student Embedded Chip and System Design Competition,
- National Second Prize, National Computer Competency Challenge (2 awards)
- National Encouragement Scholarship, Five-time Recipient
- Scholarships: First Class (2 awards), Second Class (3 awards), Third Class (2 awards)
- National Third Prize, National English Competition for College Students,
- Shanghai First Prize, National College Student Mathematics Competition,
- First-author software copyright for a robust deep-learning watermark mobile application
- Software Copyright: Code Security Analysis System Based on CodeQL and LLMs (First Author)
Services
Reviewer, NeurIPS 2026
Contributing to rigorous peer evaluation for research in machine learning and artificial intelligence.
Acknowledgements
I thank my teachers, mentors, and collaborators for their guidance and support, and my senior peers and friends for accompanying me along the way.