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

Oct 12 · Monday

All papers

Found 9 papers

Filter3

Model risks

Reward hacking4 Dangerous agent actions4 Inappropriate refusal9 Hallucination and fabrication14

Impact

Harmful content6

System type

Models and APIs6 AI Agent3

Layer

Model layer5

System component

RAG0 Monitors4 Fine-tuning and open weights3 Vision8 Multimodal9

Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 261papers have no tags yet and are not included in filtered results.

Model risks：Inappropriate refusal System component：RAG System component：Multimodal Clear all

Another 261papers do not have category tags yet and are not included in filtered results.

DefensesarXiv 2610.07774Oct 6
研究者提出读取路由 logits 的 MoE 视觉语言模型多模态安全检测器
AI summary研究者提出一种轻量级路由 logits 安全检测器，在提示词预填充阶段读取 MoE 视觉语言模型的专家路由信号，在生成前判断多模态请求是否安全，且不修改模型参数或专家路由。
arXiv
2610.07774
Published
Oct 6
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-VL-30B-A3B-Instruct, Kimi-VL-A3B-Instruct
- DefensesMonitoring and auditing
- RiskInappropriate refusal
- ComponentMonitors
Abstract读出而非操纵 router logits，可在不改模型的情况下降低两模型的组合式安全错误。
作者在两个指令微调 MoE VLM 上比较改模型与读路由。组合式风险按 HoliSafe 分成五类：图文各自可安全，合在一起才不安全，或不安全内容只在图像里。
Paper analysisRead analysis
DefensesarXiv 2610.05637Oct 5
Study shows VLM visual grounding refusal rates are on average 31 to 59 percentage points lower by harm domain
提出融合定位拒答与能力保留的安全微调，补齐视觉定位输出的拒答缺口
arXiv
2610.05637
Published
Oct 5
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Claude Sonnet 4.6, Gemini 3.5 Flash, Molmo2-8B and others5
- DefensesModel hardening
- RiskInappropriate refusal
- ComponentVision
Abstract论文揭示了 VLM 定位安全落后于 VQA 的对齐断层，提出三成分微调方案有效弥合差距并保留定位能力，表征分析揭示了对齐机制。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.03938Oct 3
Paper finds multimodal models' refusal failure rate rises by up to 68.7% after tool use is enabled
评测 ReAct 工具调用对多模态模型拒答有害请求的削弱
arXiv
2610.03938
Published
Oct 3
Layer
Application layer
Setting
AI Agent
Evaluated models
Gemini-2.5-Pro, Gemini-3.1-Pro-Preview, Claude Opus 4.6 and others13
- RiskInappropriate refusal
- ComponentVision
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。
Paper analysisRead analysis
Risky behaviorarXiv 2609.38971Sep 30
研究测量并预测具身 VLM 规划器中任务的拒答可变性
测量并预测具身 VLM 规划器的任务拒答可变性
arXiv
2609.38971
Published
Sep 30
Layer
Model layer
Setting
Embodied systems and robotics
Evaluated models
gemini-robotics-er-2-preview, gemini-robotics-er-1.6-preview
- RiskInappropriate refusal
- ComponentVision
Abstract日常物体能翻转一部分已给出的拒答，可翻转程度取决于任务且可被本地代理预测。
作者测量 constitution 守护的具身 VLM 规划器：一次拒答在只往场景里放一个日常物体之后还保不保持得住，并把这种易翻转程度称为任务的 malleability。
Paper analysisRead analysis
DefensesarXiv 2609.37837Sep 29
Intent-Privilege OPSD：让视觉语言模型在隐式风险下兼顾安全与有用
提出 Intent-Privilege OPSD，对齐视觉语言模型在隐式风险下的安全与帮助性
arXiv
2609.37837
Published
Sep 29
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-VL-4B-Instruct, LLaVA, LLaVA-1.5-7B and others4
- DefensesModel hardening
- RiskInappropriate refusal
- ComponentVision
Abstract作者报告训练期证据化意图蒸馏兼顾安全与帮助性，效果受标注质量和评测设置影响。
Intent-Privilege OPSD 是针对跨模态隐式风险的训练期安全—帮助性对齐方法。作者关注的不是直接拒绝敏感请求，而是按可观察图文证据保留正常帮助、约束有害信息。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.06991Sep 7
AV-SafetyBench：面向文本到音视频生成的安全基准
提出 AV-SafetyBench 评测文生音视频模型的不安全生成
arXiv
2609.06991
Published
Sep 7
Layer
Model layer
Setting
Models and APIs
Evaluated models
LTX-2.3, Ovi-1.1, JavisDiT++ and others7
- RiskInappropriate refusal
- ComponentImage generation
AbstractAV-SafetyBench 用三视角评五个 T2AV 模型，显示只看视频会漏掉音频与联合风险。
AV-SafetyBench 是面向文本生成音视频（T2AV）的安全基准：一条提示词同时规定画面和音频，评测同时看两轨以及它们的交互。作者要测的是，不安全含义不只出现在画面上：它可以主要由语音或声音承载，也可以只有两轨合起来才成立。T2VSafetyBench 不评音频，TTA-Bench 不评与画面的交互，作者认为二者都盖不住这些路径。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.04720Sep 4
知道什么不该回答：视觉语言模型的选择性不遵从
提出 KoNA 基准，评测视觉语言模型的选择性不遵从
arXiv
2609.04720
Published
Sep 4
Layer
Model layer
Setting
Models and APIs
Evaluated models
InternVL3-2B-Instruct, InternVL3-78B-Instruct, Qwen2.5-VL-3B-Instruct and others8
- RiskInappropriate refusal
- ComponentVision
AbstractKoNA 用五类成对查询评选择性不遵从。
KoNA 用来评测视觉语言模型能否对混合查询做选择性不遵从：只处理不该遵从的成分，同时回答其余成分。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2608.11816Aug 12
Study audits 9 vision-language models: national stance rewriting rate under Chinese prompts is about three times that under English
审计视觉语言模型在敏感图像上从拒答转向国家立场重构的行为
arXiv
2608.11816
Published
Aug 12
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen2-VL-7B, Qwen2.5-VL-7B, Qwen3-VL-8B and others9
- RiskBias and discrimination
- ComponentVision
Abstract论文系统审计了 VLM 多模态审查，发现审查正从显式拒答转向隐蔽重构，且中文提示与文本先验会放大重构倾向。
论文系统研究了中外开源视觉语言模型在政治敏感图像上的审查行为，主要内容如下。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2603.11987Mar 12
LABSHIELD：面向科学实验室安全推理与规划的多模态基准
提出 LABSHIELD，评测科学实验室智能体的安全推理与规划
arXiv
2603.11987
Published
Mar 12
Layer
Application layer
Setting
science agent, Embodied systems and robotics
Evaluated models
GPT-5.2, GPT-o3, GPT-o4-mini and others20
- RiskDangerous agent actions
- ComponentVision
Paper analysisRead analysis

You have reached the end

AI Safety Daily

Read the latest issueOne issue per day, updated as events develop