最近一年的 AI 安全论文
本站共收录 39 篇- 研究：LLM Agent 技能名幻觉率最高达 62%，可被抢注用于供应链攻击
Skills That Don't Exist: A Large-Scale Study of Hallucinated Skill Recommendation in LLM Agents
- 纹理之下的幽灵：DTS-Det 用纹理关系线索跨范式检测 AI 生成图像
Ghosts Beneath Textures: Texture-Relation Cues for Cross-Paradigm AI-Generated Image Detection
- AgentFlow：为智能体程序静态分析构建智能体依赖图
AgentFlow: Building Agent Dependency Graphs for Static Analysis of Agent Programs
- Acoda：用对抗性代码混淆防御基于大语言模型的代码分析
Acoda: Adversarial Code Obfuscation for Defending against LLM-based Analysis
- Unsafe by Flow：揭示 MCP 生态中的双向数据流风险
Unsafe by Flow: Uncovering Bidirectional Data-Flow Risks in MCP Ecosystem
- 解析与检测 GitHub Actions 中的智能体工作流注入漏洞
Demystifying and Detecting Agentic Workflow Injection Vulnerabilities in GitHub Actions
- MATRIX：基于双通道约束奇偶校验编码的多层代码水印
MATRIX: Multi-Layer Code Watermarking via Dual-Channel Constrained Parity-Check Encoding
- 评估与增强大语言模型的漏洞推理能力
Evaluating and Enhancing the Vulnerability Reasoning Capabilities of Large Language Models
- Heimdallr：高效低成本的以太坊智能合约审计智能体框架
An Effective and Cost-Efficient Agentic Framework for Ethereum Smart Contract Auditing
- 超越正确性：用多步自动定理证明揭示大语言模型推理中的逻辑缺陷
Beyond Correctness: Exposing LLM-generated Logical Flaws in Reasoning via Multi-step Automated Theorem Proving
署名以 dblp 和本人的 Google Scholar 为准；只列本站收录的 AI 安全论文。