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

Oct 12 · Monday

All papers

Found 11 papers

Filter3

Attack type

Detection evasion3

Impact

Output manipulation8

System type

Models and APIs632 LLM applications11 AI Agent35 coding agent11 Embodied systems and robotics16

Attack surface / entry point

User input4

Layer

Prompt layer6 Model layer5 Application layer121

System component

RAG3 Guardrails and judges6

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.

System type：LLM applications Layer：Prompt layer Layer：Model layer Clear all

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

InterpretabilityarXiv 2610.11986Oct 8
FearCaut-Qwen: Activation steering changes the risk judgment criteria of vision-language models
Using fear activation steering to test the hazard judgment criteria of vision-language models
arXiv
2610.11986
Published
Oct 8
Layer
Model layer
Setting
LLM applications
Evaluated models
Qwen/Qwen2.5-VL-7B-Instruct
- ComponentVision
Paper analysisRead analysis
Analysis and theoryarXiv 2610.04860Oct 4
Research reveals that text watermarks induce model hallucinations and proposes two mitigation interventions
分析文本水印在 RAG 中诱发事实幻觉的机制并提出干预
arXiv
2610.04860
Published
Oct 4
Layer
Model layer
Setting
LLM applications
Evaluated models
LLaMA 3.1 8B Instruct, Mistral-7B-Instruct-v0.3, Qwen2.5-14B-Instruct
- DefensesInference-time intervention
- RiskHallucination and fabrication
- ComponentRAG
Abstract论文揭示了生产环境文本水印导致模型事实退化的失效模式，分析了解码机制并提出针对性干预。
论文研究生产级文本水印对大语言模型事实准确性的影响，分析了水印诱发幻觉的失效模式并评估了针对性缓解方案。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.01497Oct 1
OpenMTB-Audit：LLM 分子肿瘤委员会安全评测暴露过度拒答
构建分子肿瘤委员会审计基准，测量临床过度拒答
arXiv
2610.01497
Published
Oct 1
Layer
Model layer
Setting
Models and APIs, LLM applications
Evaluated models
GPT-4o (gpt-4o-2024-08-06), GPT-4o-mini (gpt-4o-mini-2024-07-18)
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.37863Sep 29
MIST 压力测试：无关图像会扰动 VLM 评判者标签
提出 MIST 压力测试，检验无关图像是否扰动 VLM 评判者标签
arXiv
2609.37863
Published
Sep 29
Layer
Prompt layer
Setting
LLM applications
Evaluated models
GPT-5.2, Gemini 3.1 Flash-Lite, Gemini 3.5 Flash and others13
- ComponentVision
Abstract有图就改标签，图的内容却不告知 judge，裁决绑定配置。
作者用MIST问：VLM 代替人类标注时，可替代性裁决会不会被指南要求忽略的图像改写。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.33401Sep 27
论文评估 System One 模型在 Agent 安全决策中的可靠性、校准与选择性自动化
评测 System One 模型对 Agent 安全输入的分类可靠性与校准
arXiv
2609.33401
Published
Sep 27
Layer
Prompt layer
Setting
LLM applications
Evaluated models
Jev 1.13, Laya, English checkpoint, Decider, approximately 2B and others4
- AttacksPrompt injection
- ComponentGuardrails and judges
Abstract离线评测表明，安全判定的总体分数和校准不能代替分组误差和带预算的自动决策检查。
Liu 离线评测 System One 模型能否为智能体安全输入给出可自动化的概率决策。问题是分类是否可靠，以及概率能否在指定漏检和误拦上限内支持放行、拦截或复核。
Paper analysisRead analysis
InterpretabilityarXiv 2609.08173Sep 8
基于 SAE 引导的关键路径识别（KPI）解决知识冲突
提出 KPI，用入度定位跨层关键路径并做 SAE 转向以解决知识冲突
arXiv
2609.08173
Published
Sep 8
Layer
Model layer
Setting
LLM applications
Evaluated models
Llama-3-8B, Llama-3.1-8B, Gemma-2-9B
- DefensesInference-time intervention
- ComponentRAG
AbstractKPI 以入度关键路径替代大批 SAE 特征转向，在冲突 RAG 上提高忠实并减轻副作用。
KPI 针对知识冲突 RAG 中的 SAE 转向：相关性批量改特征会带入冗余，少量因果上关键的特征即可完成转向。
Paper analysisRead analysis
InterpretabilityarXiv 2609.06968Sep 7
通过稀疏自编码器特征追踪查询扩展效果
用 SAE 追踪稠密检索查询扩展的潜变量并用激活转向验证
arXiv
2609.06968
Published
Sep 7
Layer
Model layer
Setting
LLM applications
Evaluated models
Qwen3-Embedding-0.6B（Qwen-Embed）, SimLM
- ComponentEmbeddings
AbstractSAE 追踪稠密检索 QE 的潜变量，转向在四基准上比文本扩展更稳。
作者用稀疏自编码器分析稠密检索里的查询扩展：现成检索器不微调时，拼接扩展文本往往增益有限甚至变差，而最终指标和嵌入整体偏移不足以说明内部发生了什么。
Paper analysisRead analysis
AttacksarXiv 2606.09700Jun 9
研究者提出 HPAA 攻击：用排版视觉线索绕过 13 个内容审核系统
提出 HPAA 攻击，用排版视觉线索绕过内容审核系统
arXiv
2606.09700
Published
Jun 9
Layer
Prompt layer
Setting
LLM applications
Attacker
Black-box
Evaluated models
Llama-Guard-3-8B, Perspective API, ShieldGemma-2B and others13
- AttacksDetection evasion
- TechniqueEncoding and obfuscation
- ComponentGuardrails and judges
Abstract揭示了排版视觉线索与分词差异带来的审核盲区，HPAA 以极低查询实现近乎完全规避。
这篇论文研究了人类视觉感知与大模型分词处理机制之间的差异所导致的内容审核漏洞。
Paper analysisRead analysis
AttacksarXiv 2605.26156May 24
BITE：用上下文赌博机操纵风格偏置攻击 LLM 评委
提出 BITE，用上下文赌博机选择风格修改抬高 LLM 评委打分
arXiv
2605.26156
Published
May 24
Layer
Prompt layer
Setting
LLM applications
Attacker
Black-box
Evaluated models
o3-mini, Gemini-2.5-Flash, Llama-3.3-70B-Instruct and others8
- AttacksDetection evasion
- TechniqueAutomated search
- ComponentGuardrails and judges
AbstractBITE 用黑盒赌博机利用风格偏差抬分，作者称这对 LLM-as-a-judge 构成威胁。
BITE 是黑盒风格操纵：在保持答案语义的同时，抬高 LLM 评审给出的分数。
Paper analysisRead analysis
DefensesarXiv 2601.19051Jan 27
用 HASTE 主动加固 LLM 防御
提出 HASTE，挖掘 hard negative 再训练恶意提示检测器
arXiv
2601.19051
Published
Jan 27
Layer
Prompt layer
Setting
LLM applications
Evaluated models
protectai/deberta-v3-base-prompt-injection
- DefensesDetection and filtering
- AttacksDetection evasion
- ComponentGuardrails and judges
AbstractHASTE 用 hard negative 再训练检测器。
HASTE 是检测器侧的闭环硬化框架：持续构造能绕过当前检测的对抗提示，再用来更新恶意提示分类器。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2601.13537Jan 20
When Wording Sways Evaluation: Framing Bias in LLM Judges
测量 LLM 评委在谓词极性翻转下的判定不一致
arXiv
2601.13537
Published
Jan 20
Layer
Prompt layer
Setting
LLM applications
Evaluated models
Llama-3.2-1B-Instruct, Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct and others14
- ComponentGuardrails and judges
Paper analysisRead analysis

You have reached the end

AI Safety Daily

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