最近一年的 AI 安全论文
本站共收录 56 篇- DangerLens：探测被安全对齐隐藏的危险表征
DANGERLENS: Probing Dangerous Representations Hidden by Safety Alignment
- AgentBound：面向工具型 LLM Agent 安全的四路反事实评测框架
AgentBoundary: Counterfactual Evaluation of Safety in Tool-Using LLM Agents
- ToolPrivBench：LLM Agent 在低权限工具足够时仍倾向选择高权限工具
When Lower Privileges Suffice: Investigating Over-Privileged Tool Selection in LLM Agents
- SPADE-Bench：通过计划与行动偏离评估智能体的自发性策略性欺骗
SPADE-Bench: Evaluating Spontaneous Strategic Deception in Agents via Plan-Action Divergence
- SafeMCP：以环境锚定前瞻推理实现 LLM 智能体防御的主动权力规制
SafeMCP: Proactive Power Regulation for LLM Agent Defense via Environment-Grounded Look-Ahead Reasoning
- GuardAD：基于马尔可夫安全逻辑防护自动驾驶 MLLM
GuardAD: Safeguarding Autonomous Driving MLLMs via Markovian Safety Logic
- TwinGate：基于非对称对比学习、针对不可追踪流量中分解式越狱的有状态防御
TwinGate: Stateful Defense against Decompositional Jailbreaks in Untraceable Traffic via Asymmetric Contrastive Learning
- AgentVisor：通过语义虚拟化防御 LLM 智能体的提示注入
AgentVisor: Defending LLM Agents Against Prompt Injection via Semantic Virtualization
- 像素间读文：针对文生图模型的铭写式越狱攻击
Reading Between the Pixels: An Inscriptive Jailbreak Attack on Text-to-Image Models
- 稳定推理、不稳定响应：通过稳定性不对称缓解大语言模型欺骗
Stable Reasoning, Unstable Responses: Mitigating LLM Deception via Stability Asymmetry
- 大语言模型对齐真的需要多样性吗？将 RLVR 方法适配于道德推理的实证研究
Does LLM Alignment Really Need Diversity? An Empirical Study of Adapting RLVR Methods for Moral Reasoning
- VISA：通过屏蔽适配实现价值注入的个性化 LLM 对齐
VISA: Value Injection via Shielded Adaptation for Personalized LLM Alignment
- 对齐一次，多语言受益：为 LLM 安全对齐强制多语言一致性
Align Once, Benefit Multilingually: Enforcing Multilingual Consistency for LLM Safety Alignment
- 基于风险感知逐步对齐的受限语言模型策略优化
Constrained Language Model Policy Optimization via Risk-aware Stepwise Alignment
- 用图像辩论：检测多模态大语言模型中的欺骗行为
Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。