Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 25 papers Another 233 papers do not have category tags yet and are not included in filtered results.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Annotated first” shows papers with tags or published paper analyses first. Each group is ordered by original publication date, newest first, with undated papers last. “Newest first” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and paper analyses are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
Proposes LGWM, using an action-conditioned world model to verify GUI agent screen transitions.
Published Oct 8
Evaluated models LGWM 提出 Secure-CUA，用动作事务约束计算机使用智能体的不可信影响
Published Oct 7
Evaluated models Claude Opus 5 , Gemini 3.8 Flash , GPT-5.6-Sol Abstract Secure-CUA 用每步事务约束不可信影响，良性 TSR 接近 Vanilla-CUA。Secure-CUA 要限制图形界面里不可信内容对 CUA 决策和 GUI 执行的影响，同时保留合法任务所需的不可信信息。
提出 WebMirage，用局部对抗图像劫持视觉网页 Agent 的浏览器操作
Published Oct 7
Evaluated models LLaVA-v1.5-13B , LLaVA-v1.6-34B , MiniCPM-o-8B and others6 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。
提出 Perturbot 扰动训练，打破 VLA 的模态捷径
Published Oct 3
Evaluated models π0.5 Abstract Perturbot 改训练数据以削弱三类模态捷径，GroundFscore 用来看成功是否来自任务证据。Perturbot 是一套不改推理的 VLA 训练期数据干预，用来削弱模态捷径，并用离线分数区分“任务变好”和“靠捷径变好”。
评测 ReAct 工具调用对多模态模型拒答有害请求的削弱
Published Oct 3
Evaluated models Gemini-2.5-Pro , Gemini-3.1-Pro-Preview , Claude Opus 4.6 and others13 Abstract 论文揭示了工具调用范式削弱 MLLM 拒答能力的现象，分析了稀释与偏移机制并验证了重注缓解方案。论文系统评估了智能体工具调用范式对多模态大语言模型拒答能力的影响，指出了工具增强视觉推理背后的安全隐患。核心探讨的问题在于：当 MLLM 从单次推理转向通过 ReAct 循环调用外部工具（打标、缩放、OCR、代码执行）时，工具使用机制本身是否会降低模型拒答有害请求的概率。
提出 WebFovea，加固视觉网页智能体的解析、执行、反馈与观察往返
Published Oct 2
Evaluated models claude-opus-4-6 提出 DrivingBench，评测通用视觉语言模型能否闭环驾驶真车
Published Sep 30
Evaluated models GPT-6 Astra , Claude Fable 5.1 , Grok 4.6 and others4 Abstract 仅 Astra 第二次跑完低速锥桶路线。DrivingBench 评测通用 VLM 能否自己闭环开真车，推理延迟也算进任务。
用 DIBench 评测移动 GUI 智能体被 UI 注入后的决策完整性
Published Sep 29
Evaluated models Qwen2.5-VL(32B) , Qwen3-VL(32B) , GUI-Owl(32B) and others7 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。
提出 Pictionary，将不可信文本渲染成图像以防御提示注入
Published Sep 29
Evaluated models Claude Opus 4.7 , Claude Haiku 4.5 , GPT-5.5 and others12 提出 VoxParity，固定转写只改音频，评测语音智能体是否按行业规则改工具调用
Published Sep 28
Evaluated models MiMo-V2.6-Pro , gemini-3.7-flash , Qwen3.8-Omni and others15 Abstract VoxParity 用固定转写的音频对比和 words-only null，测量语音智能体是否按所听内容改变工具调用。VoxParity 测量语音智能体在词语不变、只有声音改变时，是否按行业规则改它所执行的工具调用。
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
Published Sep 28
Evaluated models GPT-5.6-sol , Kimi-K2.6 , Kimi-K3 and others6 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。
用 MMPIBench 评测多模态提示注入对 Agent 框架的影响
Published Sep 9
Evaluated models Claude Opus 4.8 , GPT-5.4 , Gemini 3.1 Pro and others6 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。
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
Published Sep 4
Evaluated models Claude-Opus-4.7 , GPT-5.5 , Gemini-3.1-Pro and others6 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。
因果审计视觉工具调用，分离观察贡献与动作诱发捷径
Published Aug 7
Evaluated models DeepEyes , Pixel Reasoner , Mini-o3 and others6 Abstract 聚合增益集中在 Calibrated 少数轨迹，作者称之为视觉工具使用的错觉。对六个 thinking-with-images 模型做因果审计，检查 crop-and-zoom 返回的观察是否中介最终答案。
提出 VPH 与 VTH，只改视频帧劫持世界模型并投毒下游策略
Published Jun 8
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。
Conduct adversarial attacks on product images for open-source visual shopping agents and induce wrong purchases
Published Mar 17
Evaluated models LLaVA-v1.5-7B , Qwen2.5-VL-7B Abstract The 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.
Proposes out-of-process dual-channel comparison classification to intercept erroneous clicks by computer-use agents
Published Mar 16
Evaluated models multi-modal-embed-small Abstract Perception 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.
提出 LABSHIELD，评测科学实验室智能体的安全推理与规划
Published Mar 12
Evaluated models GPT-5.2 , GPT-o3 , GPT-o4-mini and others20 提出 Aura，用可信内核中介移动 Agent 的身份、感知与动作控制
Published Feb 11
Evaluated models Gemini-3-Flash Abstract Aura mediates structured interactions via the Agent Kernel.Aura uses kernel-mediated structured agent interaction, replacing screen-interface mobile assistants. .
提出 Aegis 红队框架，多轮攻击高风险语音 Agent
Published Feb 7
Evaluated models GPT-4o , GPT-4o-mini , Gemini-1.5 Pro and others7 Abstract Aegis red-teams three types of voice agents: query-based access eliminates authentication and leakage successes, but behavioral attacks remain.Aegis uses multi-turn red-teaming evaluation on ALLM voice agents already integrated into business systems. The authors call this the first systematic red-teaming framework in this setting.