最近一年的 AI 安全论文
本站共收录 61 篇- REIN：通过反思与弃答对齐弥合推理与可靠性之间的差距
REIN: Bridging the Gap between Reasoning and Reliability via Reflection and Abstention Alignment
- SecureCollaRAG：抵御智能体系统知识腐化的拜占庭容错安全协同 RAG 框架
Combating Knowledge Corruption in Agent Systems: A Byzantine-Tolerant Secure Collaborative RAG Framework
- DataShield：通过共识子空间对齐识别 LLM 微调中的高风险数据
DataShield: Uncovering Risky Fine-Tuning Data Across LLMs Through Consensus Subspace Alignment
- Adaptive and Explicit Safe：触发大型推理模型的潜在安全意识
Adaptive and Explicit safe: Triggering Latent Safety Awareness in Large Reasoning Models
- TRACE：面向 LLM Agent 的任务感知自适应自进化越狱框架
TRACE: Task-Aware Adaptive Self-Evolving Agentic Jailbreaking
- 通道级语义扰动：面向多种训练范式的不可学习样本
Channel-Level Semantic Perturbations: Unlearnable Examples for Diverse Training Paradigms
- 面向大语言模型安全关键参数的识别与干预
Towards Identification and Intervention of Safety-Critical Parameters in Large Language Models
- JANUS：通过分布优化越狱文生图模型的轻量级框架
JANUS: A Lightweight Framework for Jailbreaking Text-to-Image Models via Distribution Optimization
- AttriGuard：通过工具调用的因果归因防御 LLM 智能体中的间接提示注入
AttriGuard: Defeating Indirect Prompt Injection in LLM Agents via Causal Attribution of Tool Invocations
- SpatialJB：文本分布艺术如何成为 LLM 护栏的越狱钥匙
SpatialJB: How Text Distribution Art Becomes the "Jailbreak Key" for LLM Guardrails
- MIRAGE：以黑盒且与查询无关的投毒攻击误导 RAG
MIRAGE: Misleading Retrieval-Augmented Generation via Black-box and Query-agnostic Poisoning Attacks
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。