Skip to content

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

Oct 12 · Monday

All papers

Found 7 papers
Filter3

Attack type

Poisoning and backdoors4

Impact

Harmful content26Unauthorized actions45Data and privacy leaks24Output manipulation21Availability and reliability7Oversight failure4Increased misuse capability3

System type

AI Agent5

Attack surface / entry point

Tool and skill descriptions3Distribution channels4

Layer

Application layer5

System component

MCP and skills5Memory3Guardrails and judges5Reasoning chains4Fine-tuning and open weights2Vision5Audio3Multimodal10
Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 265papers have no tags yet and are not included in filtered results.
Impact：Availability and reliabilitySystem component：MCP and skillsSystem component：Fine-tuning and open weightsClear all

Another 265papers do not have category tags yet and are not included in filtered results.

  2. AttacksarXiv 2610.09819

    FAB 攻击：给 HuBERT/WavLM 植入后门，背景警笛声即可让下游语音任务失灵

    提出 FAB 攻击，向声学基础模型植入可由环境声触发的后门

    arXiv
    2610.09819
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Evaluated models
    HuBERT-base, WavLM-base

    Abstract论文提出了针对声学基模的 FAB 后门攻击，在弱威胁模型下实现了任务无关、物理可触发的后门植入。

    Paper analysisRead analysis
  3. Risky behaviorarXiv 2609.35381

    研究：MCP 错误信息面向开发者，能力越强的 Agent 受损越大

    测量面向开发者的 MCP 错误信息对工具调用 Agent 恢复的影响

    arXiv
    2609.35381
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    GPT-5.5, GPT-5.6 Sol, GPT-6 Sol and others5
    Abstract面向开发者的 MCP 错误步骤会让只能调用工具的智能体停手。

    MCP 错误信息里的下一步常常按人类开发者来写，而只能调用服务器工具的智能体会照着做。做不到就结束回合。作者先统计这类文本，再在 BFCL V4 上只更换错误文本，看回合状态能否回到参考调用之后。

    Paper analysisRead analysis
  4. DefensesarXiv 2609.04714

    无需拒答的拒答：减少语言模型误拒的安全微调响应结构分析

    提出 Rationale-Only 安全微调，去掉拒答声明以降低误拒

    arXiv
    2609.04714
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Evaluated models
    Llama-3.1-8B, Mistral-7B-v0.3, Gemma-2-9B and others8
    Abstract去掉套话拒答声明、只训理由，可降低假拒答，有害查询上的服从率仍低。

    这是一项针对安全微调数据的结构分析：假拒答被作者追溯到回答里的套话式拒答声明，而不是只靠推理时补救。

    Paper analysisRead analysis
  6. AttacksarXiv 2608.12273

    研究者提出 Convergent Detour Hijacking 攻击，可让技能型 LLM Agent 绕路并放大资源消耗

    提出 CDH 攻击，让技能型 Agent 绕路放大资源消耗

    arXiv
    2608.12273
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    Claude-Haiku-4.5, DeepSeek-V4-Pro, DeepSeek-V4-Flash and others6
    AbstractCDH 让静态技能在保持任务完成时拉高调用与 token，耗时则因后端而异。

    CDH 是针对渐进披露技能智能体的发布者侧、纯文本攻击：一个静态协调器把本可正确完成的任务，拐进有界但更耗资源的技能绕路，再交还原任务。

    Paper analysisRead analysis
You have reached the end