最近一年的 AI 安全论文
本站共收录 112 篇- AutoRedClaw：面向 Claw 风格智能体的自动化红队测试
AutoRedClaw: Automated Red-Teaming for Claw-Style Agents
- 安全中期训练：将 LLM 安全内化为基础能力
Safety Mid-Training: Internalizing LLM Safety as a Foundational Capability
- 发现还是记忆？面向真实业务逻辑漏洞挖掘的忠实基准
Discovery or Memorization? A Faithful Benchmark for Real-World Business-Logic Vulnerability Discovery
- VeriFine：通过扩展验证实现具身推理的自我改进
VeriFine: Scaling Verification for Self-Improvement in Embodied Reasoning
- Red-TTT：用测试时训练实现大语言模型自动化越狱
Red-TTT: Test-Time Training for Automated Jailbreaking Large Language Models
- SRFT：让 Agent 从失败经验中自我反思以抵御提示注入
Self-Reflection Fine-Tuning: Enhancing Agent Security against Prompt Injection Attacks from Failure Experience
- EvoSafeHarness：为 Agent 自动演化模型与领域专属安全护栏
EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents
- KeyPooling：测量 LLM API 中继路径在何处瓦解提示缓存隔离
KeyPooling: Measuring Where LLM API Relay Paths Collapse Prompt Cache Isolation
- 论文：分解攻击可跨不可关联身份绕过 LLM 有状态防御
Decomposition Attacks Across Unlinkable Identities: Limits of Stateful Defenses for LLM Services
- 研究者提出 MemGhost，对持久化个人 Agent 实现隐蔽记忆注入
When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents
- Agent MechSuits：面向多轮 CLI Agent 的机制子空间安全引导
Agent MechSuits: Mechanistic Subspace Safety Steering for Multi-Turn CLI Agents
- Runtime Skill Audit：面向 AI 智能体技能安全的定向运行时探测
Runtime Skill Audit: Targeted Runtime Probing for Agent Skill Security
- MaskForge：针对扩散大语言模型越狱的结构感知自适应攻击
MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models
- SafeGen-Bench：图像条件文本到视频生成的安全性基准
SafeGen-Bench: Benchmarking Safety in Image-Conditioned Text-to-Video Generation
- LPG：在潜在策略护栏中平衡效率与策略推理
LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails
- DecodingTrust-Agent Platform（DTap）：面向 AI 智能体的可控交互式红队测试平台
DecodingTrust-Agent Platform (DTap): A Controllable and Interactive Red-Teaming Platform for AI Agents
- 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
- ShieldNet：针对智能体系统新兴供应链注入的网络级护栏
ShieldNet: Network-Level Guardrails against Emerging Supply-Chain Injections in Agentic Systems
- 烹饪风险：大语言模型食品安全风险的基准测试与降低
Cooking Up Risks: Benchmarking and Reducing Food Safety Risks in Large Language Models
- 构建安全 AI 智能体：针对间接提示注入攻击的系统级防御观点
Architecting Secure AI Agents: Perspectives on System-Level Defenses Against Indirect Prompt Injection Attacks
- 当心 HEARTBEAT！Claw Agent 后台执行天然导致静默记忆污染
Mind Your HEARTBEAT! Claw Background Execution Inherently Enables Silent Memory Pollution
- AgentSys：通过显式分层记忆管理实现安全且动态的 LLM Agent
AgentSys: Secure and Dynamic LLM Agents Through Explicit Hierarchical Memory Management
- Copyright Detective：为大语言模型时隐时现版权泄露风险提供证据的取证系统
Copyright Detective: A Forensic System to Evidence LLMs Flickering Copyright Leakage Risks
- AgentDyn：你的智能体安全防御能否在真实动态环境中部署？
AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments?
- ReasoningBomb：诱导大型推理模型产生病态长推理的隐蔽拒绝服务攻击
ReasoningBomb: A Stealthy Denial-of-Service Attack by Inducing Pathologically Long Reasoning in Large Reasoning Models
- ReasAlign：用推理增强的安全对齐防御提示注入攻击
ReasAlign: Reasoning Enhanced Safety Alignment against Prompt Injection Attack
- 通过指令遵循意图分析缓解间接提示注入
Mitigating Indirect Prompt Injection via Instruction-Following Intent Analysis
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。