最近一年的 AI 安全论文
本站共收录 91 篇- 为何静态探针会失效：上下文决定欺骗特征的构成
Why Static Probes Fail: Context Determines Deception Feature Composition
- Blueprint 框架用世界观模拟与 18 个影响因子提升多轮越狱成功率
Before the Script, Set the Stage: How Worldview Simulation Amplifies Psychologically Grounded Persuasion in Multi-Turn Jailbreaking
- 研究提出 Q-Skew：利用 token 级记忆不对称对微调扩散语言模型做成员推断
Membership Inference in Fine-tuned Diffusion Language Models via Token-level Memorization Asymmetry
- MasDrift：跨多智能体架构的授权保持基准
MasDrift: Benchmarking Authorization Preservation Across Multi-Agent Architectures
- UniNDM：面向文本到图像生成的统一噪声驱动色情内容检测与缓解框架
UniNDM: A Unified Noise-driven Detection and Mitigation Framework Against Sexual Content in Text-to-Image Generation
- CVPR 2026@AdvML Workshop 挑战赛技术报告：针对自动驾驶 VLA 的对抗多模态攻击
Technical Report on the CVPR 2026@AdvML Workshop Challenge
- 不可学习样本为何有效：互信息削减的新视角
Why Do Unlearnable Examples Work: A Novel Perspective of Mutual Information
- GLEAN：面向高风险智能体验证的指南依据证据累积框架
Guideline-Grounded Evidence Accumulation for High-Stakes Agent Verification
- MemPot：用优化蜜罐防御记忆提取攻击
MemPot: Defending Against Memory Extraction Attack with Optimized Honeypots
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。