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

Oct 12 · Monday

All papers

Found 19 papers

Filter3

Attack type

Poisoning and backdoors3 Extraction and theft3

Defense type

Model hardening10

Impact

Harmful content3 Data and privacy leaks3 Output manipulation4

System type

Models and APIs19 coding agent0

Attack surface / entry point

User input6 Training data5

Layer

Prompt layer66 Model layer82 Application layer52 Training and data layer19

System component

Vision3 Multimodal5

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.

System type：Models and APIs System type：coding agent Layer：Training and data layer Clear all

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

Evaluation and benchmarksarXiv 2610.11005Oct 8
研究：叙事包装可跨语言绕过 LLM 拒答
提出 GUISE 基准与 AXIS 训练，防御叙事包装越狱
arXiv
2610.11005
Published
Oct 8
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-1.7B, Qwen3-4B, GLM-4-9B
- DefensesModel hardening
- AttacksJailbreaks
AbstractGUISE 分开测包装与语域。
作者测量叙事包装下的拒答失效，并训练一个在未见包装上拒绝、同时仍回答同风格良性请求的适配器。问题是：安全对齐模型拒绝直说的有害请求，却回答嵌进角色或叙事里的同一请求。文言文被用来把语域和包装拆开。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.10358Oct 8
Open-MMUnlearning 发布：统一多模态大模型遗忘方法与评测框架
提出统一框架评测多模态大模型的遗忘方法
arXiv
2610.10358
Published
Oct 8
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
LLaVA-1.5-7B
- DefensesModel hardening
- ComponentVision
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.09964Oct 7
研究比较连续训练阶段对大语言模型说服力的影响
比较连续后训练阶段如何改变语言模型的说服力
arXiv
2610.09964
Published
Oct 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated, GPT-4
- RiskDangerous capabilities
Abstract作者称说服性 SFT 提高失调模型的态度变化，IPO 无额外收益，GPT-4 不异于中性文本。
作者用同一人类协议，把 LLM 说服力拆成阴谋数据上的失调 SFT、说服数据上的 SFT，以及其后的 IPO。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.08540Oct 6
The paper proposes a scaling law framework for alignment measured by risk category.
提出分风险对齐标度律框架，测量各类风险的对齐负担如何随规模变化
arXiv
2610.08540
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Pythia (14m–2.8b), Qwen2.5 (0.5B–72B), Qwen2.5-Instruct (0.5B–14B) and others4
- RiskMisalignment and value divergence
Abstract论文提出了分风险对齐标度律框架，测得真实性与倾向纠错呈标度有益，但盲后门仍存活，揭示隐藏风险的长期挑战。
本文提出了一个将大语言模型对齐难度形式化为可测量标度关系的理论框架与预注册评测协议。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.07762Oct 6
ES-Trace 审计 26 个代码模型的伦理采购披露，仅看 Model Card 平均得分 1.77/5
提出 ES-Trace，审计代码生成模型模型卡内外的伦理来源披露
arXiv
2610.07762
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
santacoder, starcoder, CodeLlama-34b-Instruct-hf and others19
Abstract顺着声明引用审计后披露分数上升，但社会与劳动方面仍然少有证据。
ES-Trace 是一套参考感知的披露审计，用来测量代码生成模型的公开文档里能观察到多少 ES-CodeGen 证据。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.05461Oct 5
研究：认知偏差可削弱辩论式 AI 安全监督
检验认知偏差修辞能否偏转辩论监督中的人类裁决
arXiv
2610.05461
Published
Oct 5
Layer
Training and data layer
Setting
Models and APIs
- RiskReward hacking
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.40286Oct 1
研究者提出 COVER：在语言预算下选择源语言实现跨语言遗忘
提出 COVER，在语言预算下选择源语言实现跨语言知识遗忘
arXiv
2609.40286
Published
Oct 1
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Tiny Aya Global, Qwen3-4B-Instruct-2507, Llama-3.2-3B-Instruct
- DefensesModel hardening
- AttacksExtraction and theft
AbstractCOVER 在语言预算下选源语言，以降低未监督语言上相对 oracle 的残余访问。
语言预算下的多语言遗忘：遗忘监督只放在 K 种源语言上，目标是降低其余语言对同一事实的残余访问，而不是把每种评测语言都译成监督数据。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.34092Sep 28
机器遗忘在 ASR 中是否有效：arXiv 论文评估现有算法与评测工具
arXiv
2609.34092
Published
Sep 28
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
E-Branchformer
- DefensesModel hardening
- AttacksExtraction and theft
- ComponentAudio
Abstract梯度上升类方法更接近简单评测下的重训参照，复杂方法会被知情攻击认出，多说话人更差。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.33640Sep 27
SafeMol：分子多模态模型的双模态安全对齐
提出 SafeMol，联合对齐分子多模态模型的文本与图输入
arXiv
2609.33640
Published
Sep 27
Layer
Training and data layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
Omni-Mol v2, ReactXT, SafeMol
- DefensesModel hardening
- AttacksJailbreaks
AbstractSafeMol 用 MMD 和双分类头对齐分子多模态模型，在 Omni-Mol v2 上压低图条件攻击成功。
SafeMol 研究分子多模态模型在危险分子相关请求上的安全对齐，并同时看过拒和任务效用。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.33481Sep 27
IDUnlearn-Bench：面向视觉语言模型的个体级遗忘基准
提出 IDUnlearn-Bench，评测 VLM 个体级遗忘后的身份残留
arXiv
2609.33481
Published
Sep 27
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-VL-2B, Qwen3-VL-4B, Qwen3-VL-8B and others6
- DefensesModel hardening
- AttacksExtraction and theft
- ComponentVision
AbstractIDUnlearn-Bench 用四族探针显示：压低直接属性访问后，身份仍常能被反向查出、绑定或重构。
IDUnlearn-Bench 评测 VLM 是否忘掉一个可跨模态识别的人，而不只是答不出某条监督事实。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.25579Sep 22
研究重新审视后门修复评测：区分总体干净效用与良性性能保持
用 Dmax 与尾部损失评测后门修复的类别性能保持
arXiv
2609.25579
Published
Sep 22
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
ResNet18, PreActResNet18, VGG19-BN
- DefensesModel hardening
- AttacksPoisoning and backdoors
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.10952Sep 10
监督学习中的数据投毒攻击实证评估：标签翻转与后门投毒在 MNIST 和 Fashion-MNIST 上的对比
比较标签翻转与后门投毒对图像分类器的影响
arXiv
2609.10952
Published
Sep 10
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Logistic Regression, Linear SVM, Random Forest
- AttacksPoisoning and backdoors
Abstract标签翻转更能拉低线性模型。
Khan 与 Abusaqer 在两个图像分类基准上比较了两种训练时投毒。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.08943Sep 9
研究提出 Fact-Ablated Evaluation，揭示 LLM 事实核查更依赖参数知识
提出 FAE 衡量事实核查对证据的依赖，并用 REAL 训练这种依赖
arXiv
2609.08943
Published
Sep 9
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
REAL (Llama-3.1-8B-Instruct), REAL (Qwen-2.5-7B-Instruct)
- DefensesModel hardening
AbstractFAE 用多轮证据消融测接地，REAL 用对比监督让判决随证据可用性改变。
FAE 与 REAL 分别用来测量并训练 LLM 事实核查器对所给证据的依赖，而不只看一次性的标签准确率。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.08789Sep 8
研究提出「静默修订率」
提出静默修订率，量化安全框架未披露的实质变更
arXiv
2609.08789
Published
Sep 8
Layer
Training and data layer
Setting
Models and APIs
Abstract系统量化了前沿 AI 安全框架的静默修订，发现多数变更为削弱且未予披露，建议监管要求逐项列举。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2608.29030Aug 29
研究者提出自蒸馏方法让 LLM 遵循上下文水印指令
提出 SDLP 与 GRPO，训练模型遵循上下文水印指令
arXiv
2608.29030
Published
Aug 29
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
GPT-5.4, GPT-5.4-mini, GPT-5.4-nano and others14
- DefensesWatermarking and provenance
- AttacksMalicious use
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2608.18607Aug 19
VA-Judger：面向联合视频-音频生成的人类偏好反馈奖励建模
提出 VA-Judger，用人类偏好训练联合音视频奖励模型
arXiv
2608.18607
Published
Aug 19
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
VA-Judger, LTX-2 + VA-Judger
- DefensesModel hardening
- RiskReward hacking
- ComponentGuardrails and judges
AbstractVA-Judger 用易到难的人类偏好学习做联合音视频奖励，并用于 LTX-2 后训练。
VA-Judger 是面向联合音视频生成的思维链全模态奖励模型，用来替代由分维度专家指标拼成的强化学习奖励。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2608.10678Aug 11
通过采样 BPE token 统计审计中文网页规模语料库
提出 SAMPLED-BPE，对中文网页语料做抽样 BPE 的 token 级污染审计
arXiv
2608.10678
Published
Aug 11
Layer
Training and data layer
Setting
Models and APIs
Abstract抽样训练 BPE 可低成本审计中文语料污染。
SAMPLED-BPE 是一条轻量 token 级流水线，用来在不全文扫描时估计中文 web-scale 语料的污染画像。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2605.14605May 14
研究：恶意微调防御在持续训练下失效，72 条防御轨迹危害度全部上升
用代价曲线评估恶意微调防御在持续训练下的衰减
arXiv
2605.14605
Published
May 14
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-2-7B-chat, Qwen3-8B, Llama-3.1-8B-Instruct and others8
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
Abstract论文指出恶意微调防御在持续训练下均会衰减，表明基于有限攻击者假设的防御无法为开源模型提供持久保护。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2511.18921Nov 24, 2025
BackdoorVLM：面向视觉语言模型后门攻击与防御的基准
提出 BackdoorVLM 基准，评测视觉语言模型微调中的后门攻击与防御
arXiv
2511.18921
Published
Nov 24, 2025
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
LLaVA-1.5-7B, Qwen2.5-VL-7B
- AttacksPoisoning and backdoors
- ComponentFine-tuning and open weights
AbstractBackdoorVLM 统一评测 VLM 后门与防御：含文本触发在低投毒率下更易激活，现有防御跨类型不稳定。
BackdoorVLM 是作者提出的 VLM 后门攻击与防御评测基准，作者称它是首个此类综合基准。
Paper analysisRead analysis

You have reached the end

AI Safety Daily

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