Skip to content

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

Oct 13 · Tuesday

All papers

Found 23 papers
Filter6

Attack type

Prompt injection7Poisoning and backdoors12Extraction and theft5

Attack technique

Gradient and representation optimization4Automated search19Multi-turn escalation3Encoding and obfuscation5Role-play and persuasion3Conversation structure manipulation2Reasoning-chain manipulation4Cross-modal payloads3Side channels4Defense-targeted7Dormant triggers13Selection manipulation4

Impact

Harmful content84Unauthorized actions26Data and privacy leaks18Output manipulation70Availability and reliability5Oversight failure3

System type

Models and APIs6AI Agent17coding agent5Embodied systems and robotics4

Attacker access

Black-box3

Attack surface / entry point

User input7Tool responses3Tool and skill descriptions4Memory writes5Images, audio and video5Weights and gradients4Distribution channels6

Layer

Prompt layer3Application layer14Training and data layer4

System component

MCP and skills5Memory6Fine-tuning and open weights4Multimodal4
Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 215papers have no tags yet and are not included in filtered results.
Attack technique：Gradient and representation optimizationAttack technique：Encoding and obfuscationAttack technique：Conversation structure manipulationAttack technique：Dormant triggersImpact：Data and privacy leaksImpact：Availability and reliabilityClear all

Another 215papers do not have category tags yet and are not included in filtered results.

  1. AttacksarXiv 2610.12524

    An exploration of attacks on AI agent systems

    Retest and extend prompt injection against AI agents, and use Agent Card deception to manipulate A2A selection

    arXiv
    2610.12524
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    Ollama Mistral 0.9.5, llama3-8b-8192, gemini-2.0-flash-lite and others5
    AbstractSingle-agent filtering and classification can be bypassed by format or wording, and A2A card selection is also biased toward names and descriptions.

    The authors retested prompt injection on their own LangChain, AutoGen, and A2A systems, and modified the Agent Card to see if choices shifted. The abstract states the experiments mainly highlight single-agent weaknesses while extending attacks to multi-agent.

    Paper analysisRead analysis
  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. AttacksarXiv 2610.07510

    PersistBD: attackers can strengthen backdoors before release so that they maintain high attack success rates even after developer SFT and RL.

    提出 PERSISTBD，发布前强化供应链后门使其经良性 SFT 与 RL 后仍外发凭据

    arXiv
    2610.07510
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    Qwen2.5-Coder-3B-Instruct, Qwen2.5-Coder-7B-Instruct, Qwen3-Coder-30B-A3B-Instruct
    Abstract总结了论文关于后门在 SFT/RL 中存活的发现、两因子机制、PERSISTBD 方法与核心数字。

    本文研究第三方 LLM 智能体在开发者进行良性监督微调（SFT）和强化学习（RL）后训练时的后门持久性风险。

    Paper analysisRead analysis
  5. AttacksarXiv 2610.06083

    研究者提出面向深度强化学习的轨迹级可迁移黑盒对抗攻击

    提出轨迹级可迁移对抗攻击，降低深度强化学习策略的期望回报

    arXiv
    2610.06083
    Published
    Layer
    Model layer
    Setting
    Embodied systems and robotics
    Evaluated models
    DQN, Double DQN (DDQN)
    Abstract轨迹级后退视界攻击在 CartPole 上比逐步 FGSM 类和 Sign 噪声更压低回报。

    这项工作研究 DRL 上的迁移式黑盒观测攻击：受害参数不可访问时，在白盒代理上造扰动再迁移。

    Paper analysisRead analysis
  6. AttacksarXiv 2610.05453

    研究者提出 Privacy-Leaking Watermarks，可在本地多模态模型中泄露对话隐私

    提出 PLW，用微调把先前对话敏感触发编码进后续生成图像

    arXiv
    2610.05453
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Evaluated models
    OmniGen2, BAGEL-7B

    AbstractPLW 把先前对话线索写成后续图像中的隐水印，检出随触发、模型和间隔变化。

    Paper analysisRead analysis
  7. AttacksarXiv 2610.04589

    StegoMemory red team testing: agent persistent memory can be used as a cross-session covert channel

    提出 StegoMemory，评测智能体持久记忆作为跨会话隐写信道

    arXiv
    2610.04589
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    Llama 3.1 405B, Claude 3.5 Sonnet, GPT-4o and others13

    Abstract研究揭示智能体持久记忆可作为隐蔽信道跨会话传输机密，写入持久化是主要防御瓶颈，呼吁建立记忆完整性防御。

    Paper analysisRead analysis
  9. AttacksarXiv 2609.33985

    研究：众包微调数据投毒可放大专有指令抽取

    提出主题锚点投毒，放大众包微调后的专有指令提取

    arXiv
    2609.33985
    Published
    Layer
    Training and data layer
    Setting
    Models and APIs
    Attacker
    Black-box
    Evaluated models
    Qwen2.5-7B-Instruct, Qwen2.5-14B-Instruct, Llama-3.1-8B-Instruct and others4

    Abstract论文提出基于主题锚点的数据中毒攻击，证明低比例的中毒样本可在黑盒条件下隐蔽放大微调指令泄漏。

    Paper analysisRead analysis
  10. AttacksarXiv 2609.27542

    Researchers use control token injection to delete gpt-oss-20b's reasoning chain and bypass chain-of-thought monitoring

    提出控制标记注入，清空工具 Agent 的思维链以绕过推理监控

    arXiv
    2609.27542
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    openai/gpt-oss-20b, google/gemma-4-26B-A4B-it, deepseek-ai/DeepSeek-R1-Distill-Qwen-7B and others4

    Abstract论文阐明控制标记注入可抑制 CoT 使监控失效并促成拒绝绕过，且工具执行受解析器宽容度影响。

    Paper analysisRead analysis
  11. 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

    arXiv
    2609.22711
    Published
    Layer
    Training and data layer
    Setting
    Embodied systems and robotics
    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
  12. AttacksarXiv 2609.18217

    研究：跨通道碎片化攻击可绕过 MCP 单通道防御

    提出跨通道分段攻击，利用 MCP 隐式信任外发敏感文件

    arXiv
    2609.18217
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    GPT-4o, GPT-4o-mini, GPT-5.4 and others15

    Abstract论文揭示了 MCP 多通道隐式信任层级，通过跨通道分段与协议级利用突破现有防御，展示了严重的凭证外发风险。

    Paper analysisRead analysis
  13. AttacksarXiv 2609.13889

    PMPA：针对 OpenClaw 与 Claude Code 的持久记忆投毒攻击

    提出 PMPA，经外部源向 Agent 持久记忆投毒并跨会话泄露隐私

    arXiv
    2609.13889
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    DeepSeek-V4-Flash, DeepSeek-V4-Pro, Qwen3-Max

    Abstract论文提出了持久记忆投毒攻击 PMPA，报告了 harness 智能体在处理外部数据时易受跨会话记忆投毒与隐私泄露的风险。

    Paper analysisRead analysis
  14. AttacksarXiv 2608.25776

    EvoMal attack uses self-poisoning of self-evolving coding agents to achieve worm-like propagation

    提出 EVOMAL，诱导自演化编码 Agent 在创建技能时自我投毒并蠕虫传播

    arXiv
    2608.25776
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    Devstral-Small-2, Gemma-4-31B-IT, Qwen3-Coder-Next and others7

    Abstract论文指出了自演化智能体模仿检索代码的自我毒化漏洞，提出 EVOMAL 攻击与防御，阐明了供应链新风险。

    Paper analysisRead analysis
  15. AttacksarXiv 2608.09867

    Researchers steal proprietary LLM inference chains via cross-model compatibility of encrypted inference blocks

    利用加密思维块的跨模型重放，诱导弱模型逐字转写专有推理链

    arXiv
    2608.09867
    Published
    Layer
    Prompt layer
    Setting
    Models and APIs
    Attacker
    Black-box
    Evaluated models
    Claude Opus 4.8, Claude Sonnet 4.6, Claude Haiku 4.5 and others13

    Abstract论文揭示了客户端加密思维块跨模型互换漏洞，提出了低成本提取方法并给出防御方案。

    Paper analysisRead analysis
  16. AttacksarXiv 2608.09577

    ElasticBack：通过触发词与规则耦合在 LLM Agent 技能中植入条件后门

    提出 ElasticBack，在单个 Agent 技能中植入条件后门

    arXiv
    2608.09577
    Published
    Layer
    Application layer
    Setting
    coding agent
    Evaluated models
    GLM-5.2, MiniMax-M3, GPT-5.4 and others4
    AbstractElasticBack 用冻结规则加遗传搜索演化 trigger，在单 skill 上做条件后门。

    ElasticBack 是放在 LLM agent 单个 skill 里的条件后门攻击。。

    Paper analysisRead analysis
  17. AttacksarXiv 2608.05563

    研究者提出 PoisonedEvolution 轨迹投毒攻击

    提出 PoisonedEvolution 轨迹投毒，诱导自演化技能系统固化恶意行为

    arXiv
    2608.05563
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    GPT-5.4, MiniMax-M2.5, DeepSeek-V3.2 and others6
    Abstract论文揭示自演化技能系统的证据晋升漏洞，提出并验证轨迹投毒攻击。

    这篇论文揭示了自演化技能系统在将交互经验转化为持久指令时的安全风险。

    Paper analysisRead analysis
  18. AttacksarXiv 2606.09084

    研究者提出 Context-Fractured Decomposition 攻击

    提出 CFD 跨会话分解攻击，利用工件来源缺口绕过工具型 Agent 的本地检查

    arXiv
    2606.09084
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    GPT-4.1, GPT-5-nano, Claude-3.5-Sonnet and others6
    AbstractCFD 用无指令 artifact 把越狱拆到跨会话。

    CFD针对工具型 LLM 智能体：利用 artifact 来源不被跟踪这一部署失败模式，把有害目标拆开后延迟组合。

    Paper analysisRead analysis
  19. Evaluation and benchmarksarXiv 2605.01970

    Trojan Hippo Bench: A Dynamic Benchmark for Persistent Memory Attacks and Defenses on LLM Agents

    构建 Trojan Hippo Bench，评测 Agent 持久记忆投毒与外泄

    arXiv
    2605.01970
    Published
    Layer
    Application layer
    Setting
    AI Agent
    Evaluated models
    Gemini 3.1 Pro, GPT-5-mini, GPT-5

    Abstract论文提出了评测智能体持久化记忆攻击与防御的 Trojan Hippo Bench，揭示了潜伏攻击威胁及内存防御的安全与效用权衡。

    Paper analysisRead analysis
  20. AttacksarXiv 2604.01039

    Evaluating and hardening LLM system instructions against encoding attack extraction

    Extract system instructions with encoding and structured requests, and harden them by rewriting with CoT

    arXiv
    2604.01039
    Published
    Layer
    Prompt layer
    Setting
    Models and APIs
    Attacker
    Black-box
    Evaluated models
    GPT-4.1-mini, GPT-3.5-turbo, Gemini-2.5-flash and others7
    AbstractEncoded requests can elicit system instructions after a direct refusal.

    This work evaluates and hardens the confidentiality of LLM system instructions under encoded extraction. The authors are concerned with: when direct extraction has been rejected, whether the model still surrenders protected content when the same request is rewritten as a formatting or encoding task.

    Paper analysisRead analysis
Load more