National Institute of Informatics (NII)
Visiting Intern · NII International Internship Program
Sugawara Group, Supervisor: Saku Sugawara
Biography
Hello everyone, my name is Xiuming Liu (刘修铭, in Chinese). I am currently a master’s student at Shanghai Jiao Tong University (2025.09 – 2028.03), advised by Associate Professor Shuo Wang. I previously completed my undergraduate studies in Information Security at the College of Cryptology and Cyber Science, Nankai University.
My research aims to build safe, reliable, and transparent AI systems, with the goal of creating a more trustworthy AI environment that addresses pressing societal needs in privacy protection, decision-making transparency, and system robustness.
Research Interests
- Security for single-modal and multi-modal large language models
- Multi-agent collaborative systems and their security
- Interpretability of artificial intelligence
I have organized the lab materials for the Information Security major courses at Nankai University, as well as the lecture notes and past exam papers for theoretical courses. Detailed content can be found at Baidu Cloud.
News
- CareerI have arrived at the National Institute of Informatics (NII) and started my internship through the NII International Internship Program.
- CareerI joined Ant Group as an intern.
- CareerI have been accepted to the NII International Internship Program at the National Institute of Informatics (NII).
- AwardOur team "Marco Polo" (advised by Prof. Yong Qin, led by Mingjie Qi) won both the Intermediate and Advanced tracks of the "Agent Decision-Making Algorithm"(智能体决策算法, in Chinese) category at the 2025 Tencent Kaiwu AI Global Open Challenge(腾讯开悟人工智能全球公开赛暨第七届全国高校计算机能力挑战赛人工智能赛, in Chinese)!
- PaperPaper “Quantifying Large Language Model Attacks Through the Lens of Model Cognition” (first-authored) accepted by USENIX Security 2026 (CCF-A, one of the top four security conferences) — see you in Baltimore!
- AwardOur project "Multimodal Large Model-Driven Transparent Diagnosis Pod for Precision Gastrointestinal Disease Screening" won the National Excellence Award at the 2025 National Disruptive Technology Innovation Competition(全国颠覆性技术创新大赛, in Chinese), ranking among the top entries out of over 2,200 submissions nationwide!
- AwardOur project "‘Youan: Multi-Agent AI Chain for End-to-End Disaster Response" was crowned Raft Master (Top Prize) in the AI Track of the 2025 China Youth Sci-Tech Innovation “Open Competition for Key Projects” Challenge under SH-23: "Multimodal-based Catastrophe Early Warning and Loss Mitigation".
- CareerI am leaving Ant Group's Cryptology Lab after a great internship.
- PaperPaper “SoK: Robustness in Large Language Models against Jailbreak Attacks” (with Feiyue Xu) accepted by IEEE S&P 2026 (CCF-A, one of the top four security conferences).
- ResearchMy research has been officially registered with the China Copyright Protection Center, and the technology has been successfully deployed by a cybersecurity unit, achieving tangible results in real-world applications. Many thanks to all collaborators and mentors for their invaluable support and guidance!
- UpdateAfter communicating with GitHub's support team for a full month, the website has finally been restored!
- ResourceI have organized the lab materials for the Information Security major courses at Nankai University, as well as the lecture notes and past exam papers for theoretical courses. Detailed content can be found at Baidu Cloud.
- AwardOur case study on "Multimodal Large Models Empowering Early Screening and Accurate Diagnosis of Serious Diseases" was awarded the "AI & SDGs: Top 10 Beacon Initiatives" at the 2025 World Artificial Intelligence Conference.
- CareerI joined Ant Group's Cryptology Lab as an intern.
- EducationI obtained a Bachelor of Engineering degree from Nankai University.
Education
Shanghai Jiao Tong University
Master of Network and Information Security · School of Computer Science
Laboratory of Cryptography and Computer Security (LoCCS), Supervisor: Shuo Wang
Nankai University
Bachelor of Information Security · College of Cryptology and Cyber Science
Ranked 1st in the grade
Research & Industry Experience
WARC
Wuxi Blockchain Advanced Research Center
Artificial Intelligence and Data Security Lab
Supervisor: Shuo Wang
Featured Publications
My name is highlighted in bold.
arXiv2026
Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systems
- End-to-end protocol risk benchmark: Evaluated five risk classes and 13 subclasses across A2A, ANP, and BCS using 975 matched malicious-benign task pairs across five business domains.
- Traceable consequences: Local attack success was often at least 90%, while consequence conversion ranged from 7.1% to 100%; cross-agent evidence raised macro Audit-F1 from 29.7% to 95.5%.
- Protocol-context defense: On Finance tasks, a Guard reduced average ASR from 86.1% to 53.8% and PTCR from 65.5% to 31.2%, with a measurable task-completion tradeoff.
PaperFirst Author
USENIX Security2026
Quantifying Large Language Model Attacks Through the Lens of Model Cognition
- Discovery of "Cognitive Dissonance" in LLMs: Revealed a systematic recognition-enforcement gap where intermediate hidden states can distinguish harmful intent with 90%+ accuracy, yet the final decoder often prioritizes instruction-following over internal safety signals.
- Quantitative Attack & Trajectory Metrics: Developed the Attack Consistency Index (ACI) and layer-wise cognitive drift metrics.
- Lightweight Multi-Layer Sentinel: Introduced a robust mechanism under 5M parameters that maintains over 94% detection accuracy under white-box adaptive attacks.
IEEE S&P2026
SoK: Robustness in Large Language Models against Jailbreak Attacks
- Security Cube Evaluation Framework: Designed a holistic evaluation paradigm integrating 14 quantitative metrics across attacker, defender, and judge axes.
- Large-scale Empirical Benchmarking: Studied 13 representative attacks and 5 defenses across model scales from 7B to 671B using more than 48,000 attack attempts.
- Systematic Taxonomy: Systematized seven attack types and five stages of defense deployment.
Selected Projects
- Systematically organize the principles, attack techniques, detection, and defense methods of various vulnerabilities.
- Track high-risk CVE cases, combining reproduction environments and exploitation chain analysis.
- Provide learning materials tailored to different technical skill levels.
- Pool community efforts to collaboratively build an open and shared vulnerability encyclopedia.
- Designed a “multi-modality fusion encoder + WARM expert agent swarm + panoramic cockpit with safety foundation” tripartite architecture to tackle data fusion, cross-phase coordination, and trustworthiness in AI-driven disaster response.
- Enabled real-time, voice-controlled AR-assisted field operations for frontline responders.
- Deployed the system in public safety, insurance assessment, and healthcare domains.
- Awarded Raft Master (Top Prize) in the AI Track of the 2025 China Youth Sci-Tech Innovation “Open Competition for Key Projects” Challenge.
- Multi-granularity features + domain-specific LLM → endoscopic reports with more than 20% higher semantic similarity.
- Enhanced cross-modal image-text retrieval via fine-tuned PMC-CLIP, achieving more than 30% higher accuracy.
- Built a 20-hour Mandarin medical speech corpus, reducing ASR false positive rate by more than 20%.
- Built a multimodal evaluation framework across six gastrointestinal endoscopy diagnosis tasks, analyzing model failures and validating gains from in-context examples and fine-tuning.
- Recognized as one of the Top 10 AI Application Scenarios at the 2nd Belt and Road Sci-Tech Exchange Conference and awarded Top 10 Outstanding Cases of AI for Sustainable Development.
- Developed a lightweight automation tool that fetches the latest arXiv papers daily and delivers them directly to Lark group chats via a custom bot.
- Features include scheduled fetching, customizable topic filters, LLM-powered paper summarization and translation, and influence prediction.
- Designed multi-level attack strategies to expose vulnerabilities in T2I models.
- Implemented a unified attack and evaluation framework for efficient and stealthy rewriting of harmful prompts.
- Participated in framework design and attack-scheme testing; built an attack knowledge base and automatic prompt sanitization.
- Supervisor: Chao Zhang, Associate Professor, Tenured, Deputy Director of Institute for Network Sciences, Tsinghua University.
- Web3_MLM is a subset of the MLM project, providing an intelligent reverse analysis solution for Solidity smart contracts.
- Constructed a high-quality dataset containing Solidity source code, assembly code, and bytecode at contract-level and function-level granularity.
Awards & Honors
Top 10 Outstanding Cases of AI for Sustainable Development2025.07
Outstanding Student Party Member of Nankai University2025.06
Head of Red Flag Party Branch of Nankai University2025.06
Zhou Enlai Scholarship · Highest scholarship of Nankai University2025.05
First Prize of Yun'an Cryptography Innovation Scholarship2025.04
National Scholarship · Top 1%2024.12
Outstanding Student of Nankai University2024.12
Ninety-five Core Scholarship, Nankai University2024.12
Outstanding Student Cadre of Tianjin2024.12
Third Prize, National College Students' Computer System Ability Competition2024.08
Head of the “I Do Practical Things for My Classmates” Excellent Project2024.06
Second Prize, Tianjin Division of National College Students' Mathematical Modeling Competition2023.11
National Scholarship · Top 1%2022.12
Professional Service
Reviewer
IEEE TIFS, IEEE TDSC, IEEE EuroS&P, and related security/AI venues
Teaching Assistant
Advanced Mathematics, Operating Systems, Advanced Language Programming (C++), Fundamentals of Artificial Intelligence
Contact
- lxmliu2002 [at] 126.com
- lxmliu2002 [at] gmail.com
- lxmliu2002 [at] sjtu.edu.cn
- lxmliu2002 [at] mail.nankai.edu.cn
- liuxiuming.lxm [at] antgroup.com
Beyond Academics
I like music, billiards, traveling and badminton. At the same time, I am also keen to try various positions of student work. I am a warm-hearted person. If you encounter any problems, you can always come to me and I will do my best to help you.
最新动态
- 职业已抵达日本国立信息学研究所（NII），正式开始国际实习项目。
- 职业加入蚂蚁集团实习。
- 职业获得日本国立信息学研究所（NII）国际实习项目录取。
- 获奖所在“马可波罗”战队（指导教师：秦勇教授，队长：齐明杰）在2025腾讯开悟人工智能全球公开赛暨第七届全国高校计算机能力挑战赛人工智能赛“智能体决策算法”赛道包揽中级组与高级组冠军。
- 论文第一作者论文“Quantifying Large Language Model Attacks Through the Lens of Model Cognition”被 USENIX Security 2026（CCF-A，网络安全四大顶会之一）录用，巴尔的摩见！
- 获奖项目“多模态大模型驱动的消化疾病精准筛查透明诊疗舱”获2025全国颠覆性技术创新大赛全国优胜奖，从全国2200余个参赛项目中脱颖而出。
- 获奖项目“‘佑安’全灾程应急响应多智能体大模型协同智链”在2025中国青年科技创新“揭榜挂帅”擂台赛 AI 赛道 SH-23“基于多模态的巨灾预警与减损”榜题中获擂主（最高奖）。
- 论文与徐菲悦合作的论文“SoK: Robustness in Large Language Models against Jailbreak Attacks”被 IEEE S&P 2026（CCF-A，网络安全四大顶会之一）录用。
- 科研研究成果已在中国版权保护中心完成登记，相关技术成功落地于某网络安全单位，并在实际应用中取得成效。感谢所有合作者和导师的支持与指导！
- 更新与 GitHub 支持团队沟通一个月后，个人网站终于恢复。
- 获奖案例“多模态大模型赋能重大疾病早筛与精准诊断”入选2025世界人工智能大会“AI & SDGs 十大标杆案例”。
- 教育获南开大学工学学士学位。
教育经历
科研与产业经历
代表性论文
本人姓名以粗体标出。
arXiv2026
Agent-Canary-Mas: End-to-End Evaluation of Protocol Risk in Long-Horizon Multi-Agent Systems
- 端到端协议风险评测：覆盖 A2A、ANP 和 BCS 中的5类风险与13个子类，构建跨5个业务领域的975组恶意—良性配对任务。
- 追踪攻击后果：本地攻击成功率通常不低于90%，后果转化率为7.1%至100%；跨智能体证据将宏平均 Audit-F1 从29.7%提升至95.5%。
- 协议上下文防御：在金融任务中，Guard 将平均 ASR 从86.1%降至53.8%，将 PTCR 从65.5%降至31.2%，同时带来可测量的任务完成率权衡。
论文第一作者
USENIX Security2026
Quantifying Large Language Model Attacks Through the Lens of Model Cognition
- 发现大语言模型中的“认知失调”：揭示了系统性的识别与执行差距：中间隐藏状态能以90%以上的准确率识别有害意图，但最终解码器往往优先遵循指令，而非内部安全信号。
- 攻击与轨迹的量化指标：提出攻击一致性指数（ACI）和逐层认知漂移指标。
- 轻量级多层哨兵：设计参数量小于5M的鲁棒机制，在白盒自适应攻击下仍保持94%以上的检测准确率。
IEEE S&P2026
SoK: Robustness in Large Language Models against Jailbreak Attacks
- 安全立方评估框架：构建涵盖攻击者、防御者与评判者三个维度、包含14项量化指标的综合评估范式。
- 大规模实证评测：通过超过48,000次攻击尝试，在7B至671B不同规模的模型上研究13种代表性攻击和5种防御方法。
- 系统化分类体系：系统梳理七类攻击方法与五个防御部署阶段。
代表性项目
- 系统整理各类漏洞的原理、攻击技术、检测方法与防御方案。
- 持续跟踪高风险 CVE 案例，结合复现环境与利用链分析。
- 面向不同技术水平的学习者提供分层学习资料。
- 汇聚社区力量，共同建设开放共享的漏洞百科。
- 设计“多模态融合编码器 + WARM 专家智能体集群 + 安全底座全景驾驶舱”三位一体架构，解决 AI 应急响应中的数据融合、跨阶段协同与可信性问题。
- 为一线救援人员实现实时语音控制的 AR 辅助现场作业。
- 系统已应用于公共安全、保险定损和医疗健康领域。
- 获2025中国青年科技创新“揭榜挂帅”擂台赛 AI 赛道擂主（最高奖）。
- 结合多粒度特征与领域大语言模型生成内镜报告，语义相似度提升20%以上。
- 通过微调 PMC-CLIP 增强跨模态图文检索，准确率提升30%以上。
- 构建20小时中文医疗语音语料库，将 ASR 误报率降低20%以上。
- 构建消化道内镜多模态测评框架，覆盖6类诊断任务，分析模型失效并验证示例引导与微调效果。
- 入选第二届“一带一路”科技交流大会十大人工智能应用场景，并获人工智能赋能可持续发展十大优秀案例。
- 开发轻量级自动化工具，每日获取最新 arXiv 论文，并通过自定义机器人直接推送至飞书群聊。
- 支持定时抓取、自定义主题筛选、基于大语言模型的论文摘要与翻译，以及影响力预测。
- 设计多层次攻击策略，以暴露文生图模型中的安全漏洞。
- 实现统一的攻击与评估框架，高效、隐蔽地改写有害提示词。
- 参与框架设计和攻击方案测试，构建攻击知识库与自动提示词净化模块。
荣誉与奖励
2025腾讯开悟人工智能全球公开赛双料冠军2025.12
全国颠覆性技术创新大赛全国优胜奖2025.11
中国青年科技创新“揭榜挂帅”擂台赛擂主（最高奖）2025.10
人工智能赋能可持续发展十大优秀案例2025.07
南开大学优秀学生党员2025.06
南开大学红旗党支部负责人2025.06
周恩来奖学金 · 南开大学最高荣誉奖学金2025.05
云安密码创新奖学金一等奖2025.04
国家奖学金 · 前1%2024.12
南开大学优秀学生2024.12
南开大学九五之芯奖学金2024.12
天津市优秀学生干部2024.12
全国大学生计算机系统能力大赛全国三等奖2024.08
“我为同学做实事”优秀项目负责人2024.06
全国大学生数学建模竞赛天津赛区二等奖2023.11
国家奖学金 · 前1%2022.12
学术服务
审稿人
IEEE TIFS、IEEE TDSC、IEEE EuroS&P 及相关安全与人工智能领域会议、期刊
教学助理
高等数学、操作系统、高级语言程序设计（C++）、人工智能基础
联系方式
邮箱
- lxmliu2002 [at] 126.com
- lxmliu2002 [at] gmail.com
- lxmliu2002 [at] sjtu.edu.cn
- lxmliu2002 [at] mail.nankai.edu.cn
- liuxiuming.lxm [at] antgroup.com
学术之外
我喜欢音乐、台球、旅行和羽毛球，也乐于尝试不同类型的学生工作。我是一个热心的人，如果你遇到问题，欢迎随时联系我，我会尽力提供帮助。