最近一年的 AI 安全论文
本站共收录 71 篇- 推理的一像素：一条真实关系就足以误导 LLM 智能体
The One-Pixel of Reasoning: One Truthful Relation Is Enough to Mislead LLM Agents
- 面向语言模型智能体保障的轨迹证据组织：片段、方法与残余
Organising Trajectory Evidence for Language-Model Agent Assurance: Fragments, Methods, and the Residual
- SPK：引出结构化先验知识，实现实时目标检测中可解释的分布外检测
SPK: Eliciting Structured Prior Knowledge for Interpretable Out-of-Distribution Detection in Real-Time Object Detection
- SkillFuzz：通过模糊测试技能组合来发现开放技能市场中的隐式意图
SkillFuzz: Fuzzing Skill Composition for Implicit Intents Discovery in Open Skill Marketplaces
- SCARCE：基于嵌入的可扩展级联分析用于稀有事件表征
SCARCE: Scalable Cascade Analysis for Rare-event Characterisation via Embeddings
- 深入歧义：受 A* 启发的多智能体常识混淆攻击针对 LLM 提示词
Dive into Ambiguity: A*-Inspired Multi-Agents Commonsense Obfuscation Attack on LLM Prompts
- Grounded Continuation：一种面向 LLM 对话的线性时间运行时校验器
Grounded Continuation: A Linear-Time Runtime Verifier for LLM Conversations
- FragileFlow：以谱控制约束基础模型中正确却脆弱的预测
FragileFlow: Spectral Control of Correct-but-Fragile Predictions for Foundation Model Robustness
- PrefixGuard：面向 LLM 智能体的在线失败预警与基于轨迹的诊断
PrefixGuard: Online Failure Warning and Trace-Grounded Diagnosis for LLM Agents
- S2O：用权重二阶统计量增强对抗训练
S2O: Enhancing Adversarial Training with Second-Order Statistics of Weights
- 用真相说谎：开放信道多智能体合谋借 Generative Montage 操纵信念
Lying with Truths: Open-Channel Multi-Agent Collusion for Belief Manipulation via Generative Montage
- 设计即脆弱：个性化生成中对抗性防御的局限
Fragile by Design: On the Limits of Adversarial Defenses in Personalized Generation
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。