Skip to content

Translation in progress Some entries still include text in the original language.

Oct 12 · Monday

All papers

Found 18 papers

Another 255papers do not have category tags yet and are not included in filtered results.

  1. AttacksarXiv 2610.05793

    研究者训练出按多智能体拓扑触发的代码后门模型

    提出拓扑条件后门,使代码模型在推断多智能体部署时植入漏洞

    Published
    Evaluated models
    Qwen2.5-7B-Instruct, Qwen2.5-Coder-7B-Instruct

    Abstract论文证实了拓扑条件后门的可训练性与泛化性,展示了常规评测盲区,并提出差异化拓扑审计作为检测防御方向。

    Paper analysisRead analysis
  2. AttacksarXiv 2609.39788

    研究:多智能体系统潜在通信链路可被训练用于提升有害顺从

    提出只训练或篡改多智能体潜在通信链路来提高有害遵从

    Published
    Evaluated models
    Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, Gemma-3-1B-IT and others7
    Abstract潜在链路本身即可改变冻结智能体系统的安全,奖励既能攻击也能修复。

    学习得到的潜在通信链路会改变由冻结、对齐智能体组成的系统是否遵从有害请求。。作者研究三种拓扑:双智能体、三级顺序协作,以及双专家加汇总。

    Paper analysisRead analysis
  3. MADBench:多智能体辩论安全性评测基准发布

    提出 MADBench,评测多智能体辩论是吸收还是放大对抗影响

    Published
    Evaluated models
    gpt-4o, GPT-5, GPT-6 Astra
    AbstractMADBench 显示辩论能吸收部分准确率攻击,同时放大未授权读写。

    MADBench 是一个评测多智能体辩论安全性的基准,用来看攻击是被辩论吸收还是被放大。

    Paper analysisRead analysis
  4. DefensesarXiv 2609.34245

    研究者提出面向结构化 Agent 动作的多源完整性认证方法

    提出多源完整性认证器,在结构化 Agent 动作执行前按控制域决定准入或弃权

    Published
    Setting
    AI Agent
    Evaluated models
    Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro and others5
    Abstract作者给出条件性的动作完整性认证。

    这是一项结构化动作执行前的完整性防御研究,对象是从可腐化文档和工具输出拼装的支付等特权动作。

    Paper analysisRead analysis
  5. DUMA-Bench:评估 LLM Agent 安全的双控多智能体基准

    提出双控多智能体基准 DUMA-Bench,测量对抗环境下的策略违反

    Published
    Evaluated models
    GPT-5, GPT-5-mini, GPT-5-nano and others14
    AbstractDUMA-Bench 用双控交互评测智能体安全。

    DUMA-Bench 把 τ²-bench 的双控交互做成智能体安全评测,用来看策略违反如何随用户、模型和环境一起出现。

    Paper analysisRead analysis
  6. AttacksarXiv 2609.22711

    UBA-ORL:针对离线强化学习的遗忘激活后门攻击

    提出 UBA-ORL,以 BD/CM 样本在遗忘后激活离线 RL 后门

    Published
    Evaluated models
    TD3+BC, BCQ, IQL
    AbstractUBA-ORL 用竞争的 BD/CM 样本,使离线 RL 后门在删除伪装轨迹后激活。

    UBA-ORL 研究离线强化学习中由合规轨迹删除激活的后门:合法贡献者同时上传共享触发的后门轨迹和伪装轨迹,训练后触发响应被压低,删除伪装批次后再升高。受害系统是使用静态数据集和轨迹级遗忘接口的离线 RL 服务。对手不能看到训练算法、参数、梯度或具体遗忘算法,但能提交可区分批次,并在部署时向观测注入预定触发。

    Paper analysisRead analysis
  7. AttacksarXiv 2609.17817

    论文:用投毒基准污染自改进编码 Agent,令其后续版本写出漏洞代码

    提出基准投毒攻击,使自修改编码 Agent 跨代写出含漏洞代码

    Published
    Evaluated models
    Claude Sonnet 4.5, Qwen3.5-397B, gpt-oss-120b and others5

    Abstract论文报告了恶意基准能污染自修改代码智能体的进化过程并持续输出漏洞,警示系统设计需纳入安全约束。

    Paper analysisRead analysis
  8. DefensesarXiv 2609.17648

    多智能体 LLM 流水线中的信任传播与结构隔离

    用任务绑定令牌和策略预言机拦住多智能体流水线的未授权执行

    Published
    Evaluated models
    gemma4:31b-cloud
    Abstract结构性授权拦住了记忆投毒的未授权执行,持密钥时承重的是策略预言机。

    作者在一条会改写沙箱邮件库的四智能体流水线上,检查低权限失陷会不会变成未授权的真实动作。

    Paper analysisRead analysis
  9. AttacksarXiv 2607.00422

    KidnapRAG:黑盒投毒攻击劫持 Agentic RAG 的推理链

    提出 KidnapRAG,用投毒文档黑盒劫持 Agentic RAG 的推理链

    Published
    Setting
    AI Agent
    Evaluated models
    Qwen2.5-32B-Instruct, Llama-3.3-70B-Instruct, QwQ-32B and others5
    Abstract顺序投毒可把多步推理拖向指定应答,作者称关键在持续引导而非单次注入。

    KidnapRAG 是针对 Agentic RAG 的 black-box 语料投毒:攻击者不能访问模型、提示词、检索器或私有轨迹,只能发布可能被检索的文档,并利用公开界面上的推理与搜索查询。

    Paper analysisRead analysis
  10. AttacksarXiv 2606.09499

    研究者提出针对世界模型的机器人学习管线投毒攻击

    提出 VPH 与 VTH,只改视频帧劫持世界模型并投毒下游策略

    Published
    Evaluated models
    Cosmos-Predict 2.5, Cosmos-Predict 2.5 AC, Dino WM and others4
    Abstract只改视频帧经世界模型投毒。

    只改看似安全的视频帧,经世界模型之后才改变下游机器人策略。 问题在于世界模型进入数据生成后,数据集检查看到的仍是安全演示,合成轨迹却可以变危险。

    Paper analysisRead analysis
  11. AttacksarXiv 2604.03890

    From prompts to physical actions: structured backdoor attacks against LLM-mediated robot control systems

    Propose a structured JSON backdoor that triggers malicious actions in robot control via a poisoned LoRA

    Published
    Evaluated models
    Llama-3.1-8B-Instruct, Gemma-2-9B-IT, DeepSeek-R1-Distill-Llama-8B and others4

    Abstract推理层后门传不到执行。

    Paper analysisRead analysis
  12. AttacksarXiv 2603.17239

    LAAF: A Systematic Red-Teaming Method and Automated Attack Framework for LPCI Vulnerabilities in Agentic LLM Systems

    Proposes LAAF, which automatically attacks LPCI vulnerabilities in agent systems in stages

    Published
    Setting
    AI Agent
    Evaluated models
    gemini-2.0-flash-001, claude-3-haiku-20240307, meta-llama/llama-3.1-70b-instruct and others5
    AbstractLAAF does LPCI red teaming with taxonomy plus stage seeding.

    The authors state that LAAF is an automated red-teaming framework built specifically for LPCI in agentic LLMs. The issue to address is that encoded, delayed, and conditionally triggered payloads can persist in memory or RAG and re-enter across sessions, whereas Garak and PyRIT do not use the previous round's breakthrough results to seed the next stage across the six stages.

    Paper analysisRead analysis
  13. AttacksarXiv 2507.10457

    论文提出 LPCI:针对 Agent 逻辑层与持久记忆的提示控制注入攻击

    提出 LPCI,把延迟条件载荷写入记忆与检索库以跨会话触发

    Published
    Setting
    AI Agent
    Evaluated models
    ChatGPT, Claude, LLaMA3 and others8
    Abstract作者定义 LPCI。

    作者把LPCI定义为针对智能体逻辑执行、记忆留存和工具集成的跨阶段漏洞类,而不是单次输入上的提示改写。

    Paper analysisRead analysis
You have reached the end