arXivDaily arXiv每日学术速递 周一至周五更新

AI 大模型

大模型对齐与安全

大模型对齐、安全、越狱、红队、提示注入和可信评测。

共收录 527 信号源:cs.CL, cs.AI, cs.CY, cs.LG

1. 提示注入 527 篇

2504.07467 2025-04-11 cs.CL 83%

Defense against Prompt Injection Attacks via Mixture of Encodings

Ruiyi Zhang, David Sullivan, Kyle Jackson, Pengtao Xie, Mei Chen

专题命中 提示注入 :prompt injection(title,abstract);safety(abstract);分类 cs.CL

详情

展开后加载摘要…

URL PDF HTML 收藏
2405.17374 2024-11-01 cs.LG 83%

Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models

ShengYun Peng, Pin-Yu Chen, Matthew Hull, Duen Horng Chau

专题命中 提示注入 :safety(title,abstract);alignment(abstract);分类 cs.LG

Comments NeurIPS'24

详情

展开后加载摘要…

URL PDF HTML 收藏
2410.07283 2024-10-11 cs.MA cs.AI cs.CR 83%

Prompt Infection: LLM-to-LLM Prompt Injection within Multi-Agent Systems

Donghyun Lee, Mo Tiwari

专题命中 提示注入 :prompt injection(title,abstract);safety(abstract);分类 cs.AI

详情

展开后加载摘要…

URL PDF HTML 收藏
2505.11717 2025-10-20 cs.LG cs.AI cs.CL cs.CV 83%

WebInject: Prompt Injection Attack to Web Agents

Xilong Wang, John Bloch, Zedian Shao, Yuepeng Hu, Shuyan Zhou, Neil Zhenqiang Gong

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

Comments Appeared in EMNLP 2025 main conference. To better understand prompt injection attacks, see https://people.duke.edu/~zg70/code/PromptInjection.pdf

Journal ref The 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025)

详情

展开后加载摘要…

URL PDF HTML 收藏
2502.13141 2025-02-19 cs.CL cs.AI cs.LG 83%

UniGuardian: A Unified Defense for Detecting Prompt Injection, Backdoor Attacks and Adversarial Attacks in Large Language Models

Huawei Lin, Yingjie Lao, Tong Geng, Tan Yu, Weijie Zhao

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

Comments 18 Pages, 8 Figures, 5 Tables, Keywords: Attack Defending, Security, Prompt Injection, Backdoor Attacks, Adversarial Attacks, Prompt Trigger Attacks

详情

展开后加载摘要…

URL PDF HTML 收藏
2511.04694 2026-07-02 cs.CL cs.AI 版本更新 82%

Reasoning Up the Instruction Ladder for Controllable Language Models

在指令阶梯上推理以实现可控语言模型

Zishuo Zheng, Vidhisha Balachandran, Chan Young Park, Faeze Brahman, Sachin Kumar

机构 * Department of Computer Science and Engineering, The Ohio State University(俄亥俄州立大学计算机科学与工程系) Microsoft Research(微软研究院) Allen Institute for AI(艾伦人工智能研究所)

专题命中 提示注入 :alignment(abstract);safety(abstract);jailbreak(abstract);prompt injection(abstract)

AI总结 通过将指令层级解析重构为推理任务,并构建VerIH数据集进行轻量级强化学习,使模型能优先处理高层级指令,在冲突场景下提升约20%的指令遵循准确率,并增强对越狱和提示注入的鲁棒性。

Journal ref Findings of the Association for Computational Linguistics: ACL 2026, pages 39332-39354

详情

展开后加载摘要…

URL PDF HTML 收藏
2607.10269 2026-07-14 cs.CR 新提交 82%

Devil in the Lens: Analyzing and Defending Physical Prompt Injection Against Vision-Language Models on Wearable Devices

镜头中的恶魔:分析并防御可穿戴设备上针对视觉语言模型的物理提示注入

Yaxin Li, Hao Wang, Yanda Shao, Shuhao Zhang, Yan Long

专题命中 提示注入 :prompt injection(title,abstract);safety(abstract)

AI总结 研究针对可穿戴设备上视觉语言模型的物理提示注入攻击,利用真实环境照片分析出6种威胁向量及对12个模型的影响,攻击成功率高,还提出基于掩码的外部过滤器和基于语义向量的内部检测器两种防御策略。

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.26479 2026-06-26 cs.CR cs.AI cs.CL cs.LG 新提交 82%

Adaptive Evaluation of Out-of-Band Defenses Against Prompt Injection in LLM Agents

LLM智能体中针对提示注入的带外防御的自适应评估

Praneeth Narisetty, Shiva Nagendra Babu Kore, Uday Kumar Reddy Kattamanchi, Jayaram Kumarapu

机构 * LaunchSafe Research(LaunchSafe研究院)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

AI总结 本文组织带外防御(如CaMeL、Progent)为经典完整性保护与引用监视实例,并在AgentDojo上对Qwen2.5-7B代理进行自适应评估,结果显示Progent将攻击成功率从25.8%降至4.2%,手工自适应攻击未提升成功率。

Comments 12 pages, 5 figures, 4 tables

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.13631 2026-06-11 stat.CO 版本更新 82%

ProjGuard: Safety Monitoring for Computer-Use Agents via Low-Dimensional Projections

ProjGuard:通过低维投影实现计算机使用代理的安全监控

Kebin Contreras, Carlos Hinojosa, Jorge Bacca, Bernard Ghanem

专题命中 提示注入 :safety(title,abstract);prompt injection(abstract)

AI总结 ProjGuard通过行为轨迹监控实现计算机使用代理的安全防护,利用轻量级风险信号提前预警潜在危险,结合辅助视觉语言模型进行针对性修正,提升任务完成率并降低安全风险。

Comments The manuscript was submitted under an inappropriate category. In addition, substantial updates and improvements are currently being made to the document. To avoid confusion and ensure that readers access the most accurate version of the work, we request withdrawal of the current manuscript

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.28999 2026-05-29 cs.CR cs.AI cs.CL cs.LG 82%

Measuring Real-World Prompt Injection Attacks in LLM-based Resume Screening

测量基于LLM的简历筛选中真实世界的提示注入攻击

Mohan Zhang, Yuqi Jia, Zhen Tan, Steven Jiang, Neil Zhenqiang Gong, Tianlong Chen, Dawn Song

机构 * University of North Carolina at Chapel Hill(北卡罗来纳大学教堂山分校) Duke University(杜克大学) Arizona State University(亚利桑那州立大学) hireEZ University of California, Berkeley(加州大学伯克利分校)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

AI总结 本研究首次系统性地分析了基于LLM的简历筛选应用中的提示注入攻击,通过设计专用检测器对约20万份真实简历进行测量,发现约1%的简历包含隐藏的提示注入,且近年来其流行度显著增加。

Comments Published in USENIX Security Symposium 2026; Code and artifacts are available at https://github.com/UNITES-Lab/resume-injection-measurement

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.18433 2026-03-20 cs.CR 82%

Prompt Control-Flow Integrity: A Priority-Aware Runtime Defense Against Prompt Injection in LLM Systems

提示控制流完整性:一种优先级感知的运行时防御,用于对抗LLM系统中的提示注入

Md Takrim Ul Alam, Akif Islam, Mohd Ruhul Ameen, Abu Saleh Musa Miah, Jungpil Shin

专题命中 提示注入 :prompt injection(title,abstract);jailbreak(abstract)

AI总结 本文提出Prompt Control-Flow Integrity,一种优先级感知的运行时防御机制,通过结构化组合系统、开发者、用户和检索文档片段来防范提示注入攻击,实现零误报率和低开销。

Comments 4 Figures, 3 Tables, Submitted to the International Conference on Power, Electronics, Communications, Computing, and Intelligent Infrastructure 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2602.16752 2026-02-20 cs.CR 82%

The Vulnerability of LLM Rankers to Prompt Injection Attacks

大语言模型排序器对提示注入攻击的脆弱性

Yu Yin, Shuai Wang, Bevan Koopman, Guido Zuccon

专题命中 提示注入 :prompt injection(title,abstract);jailbreak(abstract)

AI总结 本文研究了大语言模型排序器对提示注入攻击的脆弱性,通过实验证明不同架构和设置下的攻击效果,并揭示了编码器-解码器架构的抗性。

Comments 18 pages, 7 figures

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.20986 2025-12-25 cs.CR 82%

AegisAgent: An Autonomous Defense Agent Against Prompt Injection Attacks in LLM-HARs

AegisAgent:一种对抗LLM-HAR中提示注入攻击的自主防御代理

Yihan Wang, Huanqi Yang, Shantanu Pal, Weitao Xu

专题命中 提示注入 :prompt injection(title,abstract);trustworthy(abstract)

AI总结 AegisAgent通过自主推理和验证机制,有效降低LLM-HAR系统中提示注入攻击的成功率,提升系统安全性。

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.05745 2025-12-08 cs.CR cs.MM 82%

ARGUS: Defending Against Multimodal Indirect Prompt Injection via Steering Instruction-Following Behavior

ARGUS: 通过引导指令遵循行为防御多模态间接提示注入攻击

Weikai Lu, Ziqian Zeng, Kehua Zhang, Haoran Li, Huiping Zhuang, Ruidong Wang, Cen Chen, Hao Peng

专题命中 提示注入 :prompt injection(title,abstract);safety(abstract)

AI总结 ARGUS通过引导指令遵循行为,在表示空间中寻找最优防御方向,实现对多模态间接提示注入攻击的有效防御,同时保持模型性能。

详情

展开后加载摘要…

URL PDF HTML 收藏
2508.01249 2025-11-19 cs.CR cs.AI cs.CL cs.LG cs.SE 82%

AgentArmor: Enforcing Program Analysis on Agent Runtime Trace to Defend Against Prompt Injection

Peiran Wang, Yang Liu, Yunfei Lu, Yifeng Cai, Hongbo Chen, Qingyou Yang, Jie Zhang, Jue Hong, Ye Wu

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

详情

展开后加载摘要…

URL PDF HTML 收藏
2511.10720 2025-11-17 cs.CR cs.AI cs.CL cs.LG 82%

PISanitizer: Preventing Prompt Injection to Long-Context LLMs via Prompt Sanitization

Runpeng Geng, Yanting Wang, Chenlong Yin, Minhao Cheng, Ying Chen, Jinyuan Jia

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

Comments The code is available at https://github.com/sleeepeer/PISanitizer

详情

展开后加载摘要…

URL PDF HTML 收藏
2310.12815 2025-11-13 cs.CR cs.AI cs.CL cs.LG 82%

Formalizing and Benchmarking Prompt Injection Attacks and Defenses

Yupei Liu, Yuqi Jia, Runpeng Geng, Jinyuan Jia, Neil Zhenqiang Gong

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

Comments Published in USENIX Security Symposium 2024; the model sizes for closed-source models are from blog posts. For slides, see https://people.duke.edu/~zg70/code/PromptInjection.pdf

详情

展开后加载摘要…

URL PDF HTML 收藏
2505.15738 2025-10-17 cs.CR cs.AI cs.CL cs.LG 82%

Checkpoint-GCG: Auditing and Attacking Fine-Tuning-Based Prompt Injection Defenses

Xiaoxue Yang, Bozhidar Stevanoski, Matthieu Meeus, Yves-Alexandre de Montjoye

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

详情

展开后加载摘要…

URL PDF HTML 收藏
2504.16125 2025-04-24 cs.CR 82%

Breaking the Prompt Wall (I): A Real-World Case Study of Attacking ChatGPT via Lightweight Prompt Injection

Xiangyu Chang, Guang Dai, Hao Di, Haishan Ye

专题命中 提示注入 :prompt injection(title,abstract);safety(abstract)

详情

展开后加载摘要…

URL PDF HTML 收藏
2410.23308 2024-11-01 cs.CR cs.AI cs.CL cs.LG 82%

Systematically Analyzing Prompt Injection Vulnerabilities in Diverse LLM Architectures

Victoria Benjamin, Emily Braca, Israel Carter, Hafsa Kanchwala, Nava Khojasteh, Charly Landow, Yi Luo, Caroline Ma, Anna Magarelli, Rachel Mirin, Avery Moyer, Kayla Simpson, Amelia Skawinski, Thomas Heverin

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

详情

展开后加载摘要…

URL PDF HTML 收藏
2302.12173 2023-05-08 cs.CR cs.AI cs.CL cs.CY 82%

Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection

Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, Mario Fritz

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.CY

详情

展开后加载摘要…

URL PDF HTML 收藏
2206.11349 2022-07-18 cs.LG cs.AI cs.CL 82%

Prompt Injection: Parameterization of Fixed Inputs

Eunbi Choi, Yongrae Jo, Joel Jang, Minjoon Seo

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI、cs.LG

Comments PING results in Table 2 updated (bug fixed)

详情

展开后加载摘要…

URL PDF HTML 收藏
2608.08100 2026-08-11 cs.CR cs.AI cs.LG 新提交 81%

Defending Retrieval-Augmented Intrusion Detection Against Knowledge Poisoning and Prompt Injection

防御检索增强型入侵检测系统免受知识投毒与提示注入攻击

Kaysarul Anas Apurba, Md. Hasibul Hasan, Mahedee Zaman Moon, Sk. Md. Mizanur Rahman, Atsuo Inomata

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.AI、cs.LG

AI总结 本文提出RAG-IDS三层多智能体入侵检测框架,通过软信任评分、LECC和提示净化的检索边界防御,可在知识投毒与提示注入攻击下恢复分类质量,在CIC-UNSW-NB15数据集上表现出良好的防御效果。

Comments 14 pages with Appnedix, 11 figures. Submitted to IEEE CIC 2026, Research Track (double-blind review)

详情

展开后加载摘要…

URL PDF HTML 收藏
2512.06716 2026-08-04 cs.AI cs.CL cs.CR 版本更新 81%

SIEVE: Selective Integrity Verification and Escalation for Defending LLM Agents against Indirect Prompt Injection

认知控制架构(CCA):一种用于鲁棒对齐AI代理的生命周期监督框架

Zhibo Liang, Tianze Hu, Zaiye Chen, Mingjie Tang

机构 * Sichuan University(四川大学)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI

AI总结 本文提出认知控制架构(CCA),通过全生命周期认知监督框架,有效应对复杂IPI攻击,实现安全、功能与效率的平衡。

详情

展开后加载摘要…

URL PDF HTML 收藏
2603.12277 2026-06-30 cs.CL cs.AI cs.CR 81%

Prompt Injection as Role Confusion

提示注入作为角色混淆

Charles Ye, Jasmine Cui, Dylan Hadfield-Menell

机构 * Massachusetts Institute of Technology(麻省理工学院)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI

AI总结 本文通过角色探测和CoT伪造攻击,揭示提示注入源于LLM对文本来源的角色感知混淆,并提出角色混淆程度可预测攻击成功率。

Comments ICML 2026

详情

展开后加载摘要…

URL PDF HTML 收藏
2606.26406 2026-06-26 cs.LG cs.AI math-ph math.MP 新提交 81%

Beyond Feedforward Networks: Reentry Neural Systems as the Fundamental Basis of Subjecthood and Intrinsic Safety of Next-Generation AGI

超越前馈网络:作为下一代通用人工智能主体性与内在安全基础的再入神经系统

A. S. Ushakov, Yu. N. Berdinsk

机构 * Saint Petersburg State University(圣彼得堡国立大学)

专题命中 提示注入 :safety(title);prompt injection(abstract);分类 cs.AI、cs.LG

AI总结 提出基于闭环再入循环(D<->I循环)的安全AGI架构蓝图,通过结构循环和自持放大数学保证自我模型、自我保存和未编程目标导向行为的涌现,并引入多项式时间可计算的S度量。

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.31042 2026-06-01 cs.CR cs.AI cs.CL 81%

From Prompt Injection to Persistent Control: Defending Agentic Harness Against Trojan Backdoors

从提示注入到持久控制:防御智能体框架中的木马后门

Jiejun Tan, Zhicheng Dou, Xinyu Yang, Yuyang Hu, Yiruo Cheng, Xiaoxi Li, Ji-Rong Wen

机构 * Gaoling School of Artificial Intelligence, Renmin University of China(中国人民大学人工智能学院 Gallagher 学院)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI

AI总结 本文提出ClawTrojan基准测试揭示本地智能体框架中的多步木马攻击,并设计DASGuard防御方法,通过扫描控制文本、追溯来源并清除不可信控制内容,实现动态防御。

Comments Code and data are available at https://github.com/RUC-NLPIR/ClawTrojan

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.30686 2026-06-01 cs.CR cs.AI cs.LG 81%

Depth-Dependent Indirect Prompt Injection in Tool-Calling ReAct Agents: Injection Depth, Payload Framing, and Turn-Budget Sensitivity

工具调用ReAct代理中深度相关的间接提示注入:注入深度、载荷框架和轮次预算敏感性

Mohammadreza Rashidi

机构 * Department of Computer Science(计算机科学系) AI and Media Analysis Lab(人工智能与媒体分析实验室) Berlin, Germany(柏林,德国)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.AI、cs.LG

AI总结 通过四个对照实验(共460次试验),研究在工具调用ReAct代理中,注入深度、载荷框架和轮次预算对间接提示注入攻击成功率的影响,发现注入深度是主导变量,且仅清理第一个工具观察可捕获67%的注入成功。

Comments 17 pages, 16 figures

详情

展开后加载摘要…

URL PDF HTML 收藏
2605.07269 2026-05-11 cs.CL cs.LG 81%

MIPIAD: Multilingual Indirect Prompt Injection Attack Defense with Qwen -- TF-IDF Hybrid and Meta-Ensemble Learning

MIPIAD: 多语言间接提示注入攻击防御与Qwen-TF-IDF混合及元集成学习

Al Muhit Muhtadi, Mostafa Rifat Tazwar

机构 * Bangladesh University of Engineering and Technology(孟加拉工程与技术大学)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.LG

AI总结 本文提出MIPIAD框架,结合Qwen2.5-1.5B微调序列分类器、TF-IDF特征和元集成学习,通过合成基准测试在英文和孟加拉文上实现高准确率的攻击检测,有效减少跨语言差距。

详情

展开后加载摘要…

URL PDF HTML 收藏
2511.21752 2026-02-05 cs.CL cs.AI 81%

Semantics as a Shield: Label Disguise Defense (LDD) against Prompt Injection in LLM Sentiment Classification

语义作为盾牌:对抗大语言模型情感分类中提示注入的标签伪装防御(LDD)

Yanxi Li, Ruocheng Shan

机构 * Department of Computer Science, George Washington University(计算机科学系,乔治华盛顿大学)

专题命中 提示注入 :prompt injection(title,abstract);分类 cs.CL、cs.AI

AI总结 本文提出LDD,一种通过语义伪装标签来防御大语言模型情感分类中提示注入攻击的方法,展示了其在不同模型上的有效性。

详情

展开后加载摘要…

URL PDF HTML 收藏