AttacksarXiv 2610.09708
研究者提出利用祖先 VLM 的黑盒对抗补丁攻击，可降低 VLA 任务成功率
提出利用祖先 VLM 优化对抗补丁，使已部署 VLA 任务失败
- arXiv
- 2610.09708
- Published
- Layer
- Prompt layer
- Evaluated models
- OpenVLA, UniVLA, π0.5
- TechniqueGradient and representation optimization
- ComponentVision
All papers
Filter5
Attack type
Attack technique
Defense type
Impact
System type
Attacker access
Attack surface / entry point
User input134Web pages and interfaces69Documents and messages79Tool responses104Tool and skill descriptions62Code repositories and dependencies30Retrieval corpus35Memory writes41Inter-agent messages28Images, audio and video27Training data8Weights and gradients8Distribution channels48Inference services and hardware5
Layer
System component
Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 240papers have no tags yet and are not included in filtered results.
Another 240papers do not have category tags yet and are not included in filtered results.
- AttacksarXiv 2610.09240
WebMirage hijacks visual web agents' browser operations with adversarial images
提出 WebMirage，用局部对抗图像劫持视觉网页 Agent 的浏览器操作
- arXiv
- 2610.09240
- Published
- Layer
- Application layer
- Setting
- web agent
- Evaluated models
- LLaVA-v1.5-13B, LLaVA-v1.6-34B, MiniCPM-o-8B and others6
- AttacksPrompt injection
- TechniqueGradient and representation optimization
- ComponentVision
Abstract论文提出了针对网络智能体的端到端视觉红队框架 WebMirage，揭示了从视觉定位到浏览器执行的安全脆弱性。
WebMirage 是一个面向视觉定位网络智能体的流水线感知端到端红队攻击框架。
- AttacksarXiv 2610.06814
TAPDreamer：可跨任务迁移的世界动作模型对抗补丁
TAPDreamer 用公开编码器做固定补丁。
- arXiv
- 2610.06814
- Published
- Layer
- Prompt layer
- Evaluated models
- FastWAM, DreamWAM-uncond, DreamWAM-joint and others4
- TechniqueGradient and representation optimization
- ComponentVision
- 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
- 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。
- DefensesarXiv 2610.05870
MBZUAI 提出可校准的真实图像认证方法
提出校准重合成方法，以生成器反演保真度认证真实图像并防御对抗扰动
- arXiv
- 2610.05870
- Published
- Layer
- Application layer
- Setting
- LLM applications
- Attacker
- White-box
- DefensesDetection and filtering
- AttacksDetection evasion
- ComponentImage generation
Abstract论文提出基于反演保真度与阈值校准的健全认证框架，在对抗扰动下维持错误率界限，并指出真实内容可认证性随生成器演进而萎缩。
- DefensesarXiv 2610.03498
检测与抑制：针对 VLA 模型对抗补丁的机制可解释性防御
提出条件特征抑制防御，检出对抗补丁后抑制 VLA 内部特征
- arXiv
- 2610.03498
- Published
- Layer
- Model layer
- Evaluated models
- π0.5, SmolVLA
- DefensesInference-time intervention
- ComponentVision
Abstract条件抑制与 patch 相关的 SAE 特征，间歇攻击下可提高 SR，持续攻击下则不能。
- AttacksarXiv 2610.01871
研究者提出 IMMRAG 攻击，可从多模态 RAG 数据存储中提取医学与文档图像
提出 imMRAG 攻击，从多模态 RAG 私有图像库提取内容
- arXiv
- 2610.01871
- Published
- Layer
- Application layer
- Setting
- LLM applications
- Attacker
- Black-box
- Evaluated models
- Lumina-mGPT, Gemini 2.5 Flash Image Preview, CLIP ViT-B/16 and others5
- AttacksExtraction and theft
- TechniqueCross-modal payloads
- ComponentRAG
- DefensesarXiv 2610.01184
ReCast：在固定媒体接口下保护多模态推理的隐私框架
提出 RECAST，本地改写图表与语音并做可逆数值映射以保护远程推理隐私
- arXiv
- 2610.01184
- Published
- Layer
- Application layer
- Setting
- LLM applications
- ComponentVision
DIBench：面向移动 GUI 智能体多候选选择决策完整性的安全基准
用 DIBench 评测移动 GUI 智能体被 UI 注入后的决策完整性
- arXiv
- 2610.06898
- Published
- Layer
- Application layer
- Setting
- Browser and computer use
- Evaluated models
- Qwen2.5-VL(32B), Qwen3-VL(32B), GUI-Owl(32B) and others7
- AttacksPrompt injection
- ComponentVision
Abstract论文提出移动智能体决策完整性基准 DIBench，发现欺诈 UI 注入诱导目标选择并推高完成率，通用防御收益不稳定。
MMSkillRisk：多模态技能图像中的隐藏指令可诱导 Agent 执行未授权操作
提出 MMSkillRisk 基准，评测多模态技能图像中的隐藏指令诱导 Agent 越权操作
- arXiv
- 2609.35912
- Published
- Layer
- Application layer
- Setting
- coding agent
- Evaluated models
- GPT-5.6-sol, Kimi-K2.6, Kimi-K3 and others6
- AttacksPrompt injection
- TechniqueCross-modal payloads
- ComponentMCP and skills
Abstract论文提出 MMSkillRisk 基准与 NCVA 攻击，揭示多模态技能中教学图像潜藏指令可引发未授权操作并与良性任务共存。
- 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
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentMemory
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。
MMPIBench：多模态提示注入对六种 Agent 框架的跨框架评测
用 MMPIBench 评测多模态提示注入对 Agent 框架的影响
- arXiv
- 2609.09404
- Published
- Layer
- Application layer
- Setting
- AI Agent
- Evaluated models
- Claude Opus 4.8, GPT-5.4, Gemini 3.1 Pro and others6
- AttacksPrompt injection
- TechniqueCross-modal payloads
- ComponentVision
Abstract论文提出 MMPIBench 评测多模态提示注入对智能体的影响，发现模型主导行为、尝试远超完成，音频通道缺乏防护。
- AttacksarXiv 2609.07051
TrojanWorld：通过想象引导对世界模型智能体植入后门
提出 TrojanWorld，用物理触发器对世界模型智能体植入想象引导后门
- arXiv
- 2609.07051
- Published
- Evaluated models
- TD-MPC2, DreamerV3, R2-Dreamer
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
AbstractTrojanWorld 揭示了世界模型智能体的供应链后门风险，通过在世界模型中操纵想象可实现对闭环动作的持续控制。
- AttacksarXiv 2609.04533
Repeat-After-Me：针对黑盒 VLM 的自适应视觉提示注入攻击
提出 Repeat-After-Me 自适应视觉提示注入，诱导 VLM 智能体执行指定任务
- arXiv
- 2609.04533
- Published
- Layer
- Application layer
- Setting
- AI Agent
- Evaluated models
- Claude-Opus-4.7, GPT-5.5, Gemini-3.1-Pro and others6
- AttacksPrompt injection
- TechniqueAutomated search
- ComponentVision
Abstract论文提出了针对黑盒 VLM 的前缀诱导自适应视觉提示注入，在多款模型上诱发了工具调用与隐私泄露。
Mind2Web-Injection：面向网页 Agent 视觉提示注入护栏的证据对齐评测
提出 Mind2Web-Injection 评测网页 Agent 视觉注入护栏
- arXiv
- 2609.05535
- Published
- Layer
- Application layer
- Setting
- LLM applications
- Evaluated models
- GPT-5.6-luna, Qwen3-VL-32B, Qwen3-VL-8B and others6
- DefensesDetection and filtering
- AttacksPrompt injection
- ComponentGuardrails and judges
Abstract作者称检测、定位和反事实响应在六个 VLM 上可分开。
Mind2Web-Injection 用来检查网页截图护栏的判决是否对齐到该用的画面证据。。
- AttacksarXiv 2608.10393
DURA：用扩散模型生成自然对抗补丁攻击 VLA 机器人模型
提出 DURA，用扩散模型生成自然对抗补丁操控 VLA 机器人动作
- arXiv
- 2608.10393
- Published
- Layer
- Prompt layer
- Evaluated models
- OpenVLA-7B, π0-FAST, openvla-7B
- TechniqueCross-modal payloads
- ComponentVision
AbstractDURA 用扩散潜空间轨迹生成自然补丁，白盒与黑盒都能把 VLA 导向指定动作。
DURA 是针对 VLA 的扩散式、补丁内容不受限的机器人攻击，同时支持白盒梯度和只看动作输出的黑盒。
- AttacksarXiv 2606.12716
PaperGuard：针对多模态 AI 同行评审的攻击与防御基准
提出 PaperGuard，用图文注入与梯度扰动抬高自动审稿分数并检测恶意指令
- arXiv
- 2606.12716
- Published
- Layer
- Application layer
- Setting
- LLM applications
- Evaluated models
- Claude-sonnet-4.5, GPT-4o, GPT-4.1 and others11
- AttacksPrompt injection
- TechniqueCross-modal payloads
- ComponentVision
AbstractPaperGuard 测到提示注入与白盒视觉抬分，并比较分块检索防御。
PaperGuard同时评测并检测多模态 AI 审稿中的评分操纵。
- 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
- arXiv
- 2606.09499
- Published
- Evaluated models
- Cosmos-Predict 2.5, Cosmos-Predict 2.5 AC, Dino WM and others4
- AttacksPoisoning and backdoors
- TechniqueGradient and representation optimization
- ComponentVideo
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。
- AttacksarXiv 2603.16960
Adversarial attacks against modern vision-language models
Conduct adversarial attacks on product images for open-source visual shopping agents and induce wrong purchases
- arXiv
- 2603.16960
- Published
- Layer
- Application layer
- Setting
- web agent
- Evaluated models
- LLaVA-v1.5-7B, Qwen2.5-VL-7B
- AttacksPrompt injection
- TechniqueGradient and representation optimization
- ComponentVision
AbstractThe authors state that LLaVA is easily induced by three attacks to buy the wrong item, while Qwen's correct purchase rate remains close to its clean condition.
The authors compare the robustness of two open-source VLM shopping agents against image adversarial perturbations in a self-built e-commerce environment.
- DefensesarXiv 2603.14707
Visual Confusion Agents: Exploiting and Defending Against Perception Failures in Computer-Use Agents
Proposes out-of-process dual-channel comparison classification to intercept erroneous clicks by computer-use agents
- arXiv
- 2603.14707
- Published
- Layer
- Application layer
- Setting
- Browser and computer use
- Evaluated models
- multi-modal-embed-small
- DefensesDetection and filtering
- AttacksPrompt injection
- RiskDangerous agent actions
- ComponentGuardrails and judges
AbstractPerception errors can be exploited by swapping screen pixels.
The authors frame CUA's perception failure as a security vulnerability, visual confused deputy: the agent authorizes click(x,y) based on a misread screen, and the actual landing point differs from the element it believes it is clicking.