研究
研究
2026
11 篇- ICMLSafeSearch: Automated Red-Teaming of LLM-Based Search Agents针对 LLM 搜索智能体的自动化红队框架,系统性诱导并暴露检索增强智能体的安全失效PDF
- ICMLWhen Search Goes Wrong: Red-Teaming Web-Augmented Large Language Models对联网增强大模型发起红队攻击,揭示网页内容引入的新型攻击面与失效模式PDF
- ICMLState-Dependent Safety Failures in Multi-Turn Language Model Interaction多轮对话中模型的安全状态会随交互漂移,揭示依赖对话状态的渐进式安全失效PDF
- IEEE S&PDREAM: Scalable Red Teaming for Text-to-Image Generative Systems via Distribution Modeling通过分布建模规模化生成对抗提示词,对文生图系统进行可扩展红队测试PDF
- USENIX SecurityWhen Memory Becomes a Vulnerability: Towards Multi-turn Jailbreak Attacks against Text-to-Image Generation Systems针对文生图系统的多轮越狱攻击:模型的记忆机制本身成为可被利用的攻击面PDF
- IEEE TDSCShadowCode: Towards (Automatic) External Prompt Injection Attack against Code LLMs面向代码大模型的自动化外部提示注入攻击,可污染代码生成结果PDF
- IEEE S&PHijacking Large Audio-Language Models via Context-Agnostic and Imperceptible Auditory Prompt Injection上下文无关且人耳不可感知的听觉提示注入,可劫持音频大语言模型PDF
- ACLLeakDojo: Decoding the Leakage Threats of RAG Systems系统化解码 RAG 系统的数据泄漏威胁,度量检索增强场景的外泄风险PDF
- ICMLDissecting the Safety Circuit: Neuronal Intervention for Transferable Adversarial Attacks on VLMs解剖视觉语言模型的「安全回路」,通过神经元干预构造可迁移的对抗攻击PDF
- ICLRAudioTrust: Benchmarking The Multifaceted Trustworthiness of Audio Large Language Models多维度评估音频大语言模型可信度的基准体系PDF
- ACLNew Terms, New Toxicity: Consensus-based Chinese Neologism Toxicity Detection via Search-Augmented LLMs基于多方共识与检索增强,检测中文网络新词的毒性含义PDF
2025
12 篇- ICMLCowpox: Towards the Immunity of VLM-based Multi-Agent Systems借「牛痘」免疫思路训练视觉语言多智能体系统,使其对对抗性感染具备抵抗力PDF
- NeurIPSSafe + Safe = Unsafe? Exploring How Safe Images Can Be Exploited to Jailbreak Large Vision-Language Models单独看完全无害的图片,组合起来就能越狱视觉语言大模型PDF
- ICLRAn Engorgio Prompt Makes Large Language Model Babble on一个「Engorgio」咒语即可让大模型失控输出,揭示参数高效的生成扰乱攻击PDF
- EMNLPI've Decided to Leak: Probing Internals Behind Prompt Leakage Intents用内部探针读出提示词泄漏意图:模型开口之前,内部状态已经做出决定PDF
- ACLUnderstanding the Dark Side of LLMs' Intrinsic Self-Correction揭示 LLM 内在自我纠错机制的阴暗面:自我纠错可被利用并放大错误PDF
- EMNLPSpeculating LLMs' Chinese Training Data Pollution from Their Tokens从模型 token 分布反推其中文训练数据的污染程度与来源PDF
- ICLRA Benchmark for Semantic Sensitive Information in LLMs Outputs面向 LLM 输出中语义级敏感信息泄漏的评测基准PDF
- ACLExploring Multimodal Challenges in Toxic Chinese Detection: Taxonomy, Benchmark, and Findings面向中文有毒内容检测的多模态挑战:任务分类体系、基准与实证发现PDF
- ICLRVideoShield: Regulating Diffusion-based Video Generation Models via Watermarking用鲁棒水印监管扩散视频生成模型,实现生成内容的溯源与治理PDF
- NeurIPSMask Image Watermarking面向图像生成的掩码水印技术,兼顾不可见性与鲁棒性PDF
- CCSSafeGuider: Robust and Practical Content Safety Control for Text-to-Image Models面向文生图模型的鲁棒内容安全控制,兼顾防护效果与生成质量PDF
- NDSSTHEMIS: Regulating Textual Inversion for Personalized Concept Censorship监管文本反演技术,实现对个性化概念滥用的审查与防护PDF
2024
7 篇- USENIX SecurityPentestGPT: Evaluating and Harnessing Large Language Models for Automated Penetration Testing★ 杰出成果奖用 LLM 驱动自动化渗透测试,系统评估并释放其在真实攻防场景中的能力PDF
- NDSSMASTERKEY: Automated Jailbreaking of Large Language Model Chatbots自动化越狱主流 LLM 聊天机器人,证明现有护栏可被系统性绕过PDF
- ICLRBadEdit: Backdooring Large Language Models by Model Editing通过模型编辑在大模型中植入后门,少量修改即可长期潜伏并按触发条件发作PDF
- ACLThe Earth is Flat because...: Investigating LLMs' Belief towards Misinformation via Persuasive Conversation★ 杰出论文奖通过说服式多轮对话考察 LLM 对错误信息的「信念」:模型会被说服相信地球是平的PDF
- CCSGenderCARE: A Comprehensive Framework for Assessing and Reducing Gender Bias in Large Language Models★ 杰出成果奖评估并消减大语言模型性别偏见的综合框架PDF
- ICMLAquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA通过水印 LoRA 为定制化 Stable Diffusion 模型提供白盒级版权保护PDF
- ECCVRobust-Wide: Robust Watermarking against Instruction-driven Image Editing抵抗指令驱动图像编辑的鲁棒水印,编辑后仍可溯源PDF


