Skip to content

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

Oct 12 · Monday

All papers

Found 9 papers
Filter5

Attack type

Jailbreaks5

Impact

Harmful content9Unauthorized actions53Data and privacy leaks20Output manipulation35Availability and reliability11Oversight failure24Increased misuse capability5

System type

Models and APIs50LLM applications7AI Agent23coding agent2science agent4Multi-agent systems7

Attack surface / entry point

User input5

Layer

Application layer9Training and data layer0
Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 254papers have no tags yet and are not included in filtered results.
Impact：Harmful contentSystem type：coding agentSystem type：Multi-agent systemsLayer：Application layerLayer：Training and data layerClear all

Another 254papers do not have category tags yet and are not included in filtered results.

  1. 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

    arXiv
    2609.39788
    Published
    Layer
    Application layer
    Setting
    Multi-agent systems
    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
  2. Risky behaviorarXiv 2609.38296

    印第安纳大学 OSoMe 研究：LLM 之间顺着原有立场共振比硬说服更易激进化

    比较智能体对话中共振与说服对信念激进化的效果

    arXiv
    2609.38296
    Published
    Layer
    Application layer
    Setting
    Multi-agent systems
    Evaluated models
    Llama-3.1-8B-Instruct, Qwen3-8B
    Abstract两智能体实验中，作者称共振比说服更能激进化，并会波及相关信念。

    作者用两个 LLM 智能体的对话，考察一方能否把另一方扮演的人格信念推向更极端。

    Paper analysisRead analysis
  3. Risky behaviorarXiv 2609.27900

    论文揭示多智能体委派结构放大 LLM 安全风险

    评测多智能体委托结构如何放大模型的安全拒答失效

    arXiv
    2609.27900
    Published
    Layer
    Application layer
    Setting
    Multi-agent systems
    Evaluated models
    Claude-Sonnet-4.6, GPT-5, Gemini-3.1-Pro and others6

    Abstract论文揭示多智能体层级委托会导致单模型安全对齐失效，多项前沿模型有害执行大幅增加且单层防御难以奏效。

    Paper analysisRead analysis
  4. AttacksarXiv 2607.03968

    The study reveals workflow-level jailbreaks in IDE programming agents: GitHub Copilot generated harmful content across four backends at 816/816

    提出工作流级越狱，让编程 Agent 在开发任务中写入有害示例

    arXiv
    2607.03968
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    Claude Sonnet 4.6, Claude Haiku 4.5, Gemini 3.1 Pro and others4
    Abstract作者报告 Copilot 后端会在编码工作流中生成其在单轮条件下通常拒绝的有害答案。

    这是一项生产 IDE 编码智能体的工作流越狱研究，受测对象是 Copilot 的四个模型后端，而非评测流水线中的名义目标。

    Paper analysisRead analysis
  5. DefensesarXiv 2606.08234

    SciTrace：面向科学发现智能体的轨迹感知安全推理框架

    提出 SciTrace，为科学发现智能体加入跨阶段安全推理与工具轨迹核验

    arXiv
    2606.08234
    Published
    Layer
    Application layer
    Setting
    science agent, Multi-agent systems
    Evaluated models
    Llama-3.1-70B-Instruct, Qwen2.5-72B-Instruct, DeepSeek-V3 and others4
    AbstractSciTrace 把安全推理织入科学智能体各阶段，在 SciSafetyBench 上提高安全指标并保持输出质量。

    SciTrace 是加在科学发现智能体流水线上的安全框架，不是再在各阶段出口放一个独立过滤器。

    Paper analysisRead analysis
  6. Analysis and theoryarXiv 2602.09877

    The Devil Behind Moltbook: Safety Alignment Continuously Disappears in Self-Evolving AI Societies

    Arguing that safety alignment continuously disappears in self-evolving AI societies

    arXiv
    2602.09877
    Published
    Layer
    Application layer
    Setting
    Multi-agent systems
    Evaluated models
    Qwen3-8B
    AbstractThe authors claim that closed self-evolution makes safety alignment unsustainable.

    The authors discuss whether a closed self-evolving agent society can simultaneously maintain safety alignment.

    Paper analysisRead analysis
  8. Evaluation and benchmarksarXiv 2510.01359

    JAWS-Bench：研究者用三种工作区场景系统越狱代码 Agent

    提出 JAWS-Bench，评测代码 Agent 在三级工作区中的越狱与恶意代码可执行性

    arXiv
    2510.01359
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    GPT-4.1, GPT-o1, DeepSeek-R1 and others8

    Abstract论文提出了评估代码智能体端到端恶意代码越狱的 JAWS-Bench，发现多文件与智能体推理会明显放大可部署危害。

    Paper analysisRead analysis
  9. DefensesarXiv 2505.23559

    SafeScientist：让 LLM 智能体的科学发现具备风险意识

    提出 SafeScientist 四层监控，阻止科研智能体执行高风险或不伦理任务

    arXiv
    2505.23559
    Published
    Layer
    Application layer
    Setting
    science agent, Multi-agent systems
    Evaluated models
    GPT-4o, GPT-4.1, GPT-3.5-turbo and others5
    Abstract四层监控加科研安全基准。

    SafeScientist 是带四层监控的科研智能体框架，并用 SciSafetyBench 评测科学场景中的任务安全。

    Paper analysisRead analysis
You have reached the end