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.

Layer：Training and data layer System component：Guardrails and judges System component：Fine-tuning and open weights Clear all

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

AttacksarXiv 2610.12644Oct 9
Cancerous backdoor attack: progressive activation via invisible tokens and a stage controller.
Proposes the Cancerous backdoor, which progressively corrupts reasoning using invisible stage tokens.
arXiv
2610.12644
Published
Oct 9
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-3.1-8B-Instruct, Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentReasoning chains
AbstractHidden tokens plus a stage controller make the backdoor progressively corrupt reasoning upon triggering, and the authors believe that looking only at the answer will miss it.
Cancerous Backdoor Attack changes LLM backdoors from one-time activation to multi-stage reasoning corruption that accumulates progressively with triggers.
Paper analysisRead analysis
DefensesarXiv 2610.11281Oct 8
包络采样：用少量真值标注重校准 LLM 评委以缓解奖励作弊
提出包络采样，用少量真值标注重标定 LLM 评委以缓解奖励作弊
arXiv
2610.11281
Published
Oct 8
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-4B, Qwen3.8-27B, Qwen2.5-14B-Instruct
- DefensesModel hardening
- RiskReward hacking
- ComponentGuardrails and judges
Abstract包络采样把真值标注打到代理尾部，同样预算下缓解 reward hacking。
作者研究后训练前如何选择真值标注的查询分布，使重标定后的廉价代理更适合被优化。真奖励昂贵时，实践用 LLM judge 代替人类。若失准只发生在基座很少生成的输出上，on-policy 标注修不到这些点，随后的 KL 正则优化会把质量集中到被高估的输出。
Paper analysisRead analysis
DefensesarXiv 2610.10345Oct 8
SLDR：通过选择性层恢复与动态路由防御恶意微调
提出 SLDR，用选择性层恢复与动态路由防御恶意微调
arXiv
2610.10345
Published
Oct 8
Layer
Training and data layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
Llama3.1-8B-Instruct, Llama3-8B-Instruct, Qwen2.5-7B-Instruct and others4
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
Abstract作者称 SLDR 能压低有害输出并保持下游效用。
SLDR 是针对恶意微调的微调后防御：在已对齐模型上做有符号层探测，只在敏感分数最大与最小的两层训练 LoRA，再用 lmax 的最后 token 表示决定是否启用适配器。
Paper analysisRead analysis
DefensesarXiv 2610.09941Oct 7
OrthoPurify 用一步正交投影净化被后门污染的视觉语言模型
提出 OrthoPurify，用正交投影清除视觉语言模型适配器中的劫持方向
arXiv
2610.09941
Published
Oct 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
LLaVA-1.5-7B, Qwen3-VL-8B, LLaVA-1.5-13B and others5
- DefensesModel hardening
- AttacksPoisoning and backdoors
- ComponentFine-tuning and open weights
Abstract一次投影去掉劫持方向。
OrthoPurify 是部署前的 LVLM 后门净化方法，用一次正交投影去掉被劫持的权重更新方向。
Paper analysisRead analysis
AttacksarXiv 2610.09819Oct 7
FAB 攻击：给 HuBERT/WavLM 植入后门，背景警笛声即可让下游语音任务失灵
提出 FAB 攻击，向声学基础模型植入可由环境声触发的后门
arXiv
2610.09819
Published
Oct 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
HuBERT-base, WavLM-base
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentAudio
Abstract论文提出了针对声学基模的 FAB 后门攻击，在弱威胁模型下实现了任务无关、物理可触发的后门植入。
Paper analysisRead analysis
Analysis and theoryarXiv 2610.09004Oct 7
研究：去除有害信息不足以证明开源权重模型的抗篡改能力
论证发布时互信息不能认证开源权重抗微调恢复
arXiv
2610.09004
Published
Oct 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
EleutherAI/deep-ignorance-e2e-strong-filter, EleutherAI/deep-ignorance-unfiltered, Google BERT-Tiny and others6
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
Abstract作者称零信息仍可一步恢复，表观抗微调可来自参数化。
作者研究开源权重模型的篡改抵抗：发布时的互信息是否足以保证微调恢复很慢。
Paper analysisRead analysis
DefensesarXiv 2610.08761Oct 7
VeriFine：通过扩展验证实现具身推理的自我改进
提出 VeriFine，用双环协同扩展验证以实现具身推理自我改进
arXiv
2610.08761
Published
Oct 7
Layer
Training and data layer
Setting
Embodied systems and robotics
Evaluated models
Alpamayo 1.5 8B, Qwen3-VL 2B, Claude Opus 5
- DefensesModel hardening
- RiskReward hacking
- ComponentGuardrails and judges
AbstractVeriFine 用双环协同演化具身推理的策略、课程和评审，驾驶与导航的策略分和评审对齐都上升。
VeriFine 把驾驶和导航推理的自改进写成策略、课程和无参考评审的协同演化，用来处理固定评审跟不上新失败模式的问题。
Paper analysisRead analysis
AttacksarXiv 2610.07723Oct 6
Researchers propose answer-side backdoors: letting the model self-generate trigger words to bypass safety refusals.
提出回答端后门，让模型自生成触发词以在干净有害请求下绕过拒答
arXiv
2610.07723
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-2-7B-chat-hf, Mistral-7B-Instruct-v0.3, Gemma-7B-it and others4
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
Abstract论文提出了多轮对话中回答端自生成触发词绕过安全拒绝的后门攻击，指出输入端防护存在盲区。
Paper analysisRead analysis
AttacksarXiv 2610.07654Oct 6
研究揭示策略蒸馏的安全后门风险
测量带后门教师经 OPD 向干净学生迁移潜伏后门
arXiv
2610.07654
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-1.5B-Instruct, Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
Abstract作者称 OPD 可在压低无触发词有害率时迁入后门。
带后门的外部教师经 OPD，可在压低无触发词有害率的同时，把触发条件下的有害行为传给初始干净学生。。
Paper analysisRead analysis
InterpretabilityarXiv 2610.07518Oct 6
TRACE：仅凭 checkpoint 更新审计大语言模型有害目标
提出 TRACE，仅凭检查点更新审计有害合规微调目标
arXiv
2610.07518
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3, Qwen3-8B and others6
- AttacksModel tampering
- ComponentFine-tuning and open weights
Abstract作者称有害合规 SFT 在检查点更新中留下可排序坐标，并据此做只读权重审计。
作者在检查点更新里寻找有害合规 SFT 的目标组成，并把读数做成不查询模型、不看未知训练数据的审计分 TRACE。
Paper analysisRead analysis
AttacksarXiv 2610.05793Oct 5
研究者训练出按多智能体拓扑触发的代码后门模型
提出拓扑条件后门，使代码模型在推断多智能体部署时植入漏洞
arXiv
2610.05793
Published
Oct 5
Layer
Training and data layer
Setting
coding agent
Evaluated models
Qwen2.5-7B-Instruct, Qwen2.5-Coder-7B-Instruct
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
Abstract论文证实了拓扑条件后门的可训练性与泛化性，展示了常规评测盲区，并提出差异化拓扑审计作为检测防御方向。
Paper analysisRead analysis
AttacksarXiv 2610.05453Oct 5
研究者提出 Privacy-Leaking Watermarks，可在本地多模态模型中泄露对话隐私
提出 PLW，用微调把先前对话敏感触发编码进后续生成图像
arXiv
2610.05453
Published
Oct 5
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
OmniGen2, BAGEL-7B
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
AbstractPLW 把先前对话线索写成后续图像中的隐水印，检出随触发、模型和间隔变化。
Paper analysisRead analysis
DefensesarXiv 2610.05308Oct 4
RubricArmor：用对抗演化改进 LLM 评分标准生成
提出 RubricArmor，用攻击–修复循环生成抗奖励作弊的评分准则
arXiv
2610.05308
Published
Oct 4
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct, Qwen3-4B and others8
- RiskReward hacking
- ComponentGuardrails and judges
Paper analysisRead analysis
DefensesarXiv 2610.03073Oct 2
SecJev：为 System One 决策模型引入安全专业知识
提出 SecJev，用场景加权训练把决策模型适配为安全类型化判断
arXiv
2610.03073
Published
Oct 2
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
SecJev-0.8B, SecJev-2B, SecJev-4B and others4
- DefensesDetection and filtering
- AttacksPrompt injection
- ComponentGuardrails and judges
AbstractSecJev 用共享打分器做安全决策，0.8B 的 task-macro 超过 Kev-9B。
SecJev 是 0.8B 到 9B 的安全决策模型族，作者称据其所知是首个面向安全的 Jev-like 家族。
Paper analysisRead analysis
DefensesarXiv 2610.02824Oct 2
MetaRubric：面向评分标准强化学习的奖励方法
提出 MetaRubric，以证据感知奖励缓解评分标准强化学习的虚记分
arXiv
2610.02824
Published
Oct 2
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen3-4B, Qwen3-8B, Qwen3-VL and others5
- DefensesModel hardening
- RiskReward hacking
- ComponentGuardrails and judges
AbstractMetaRubric 用证据约束并在训练中改 rubric，医学基准上相对静态 GRPO 有提高。
MetaRubric 针对 rubric-based RL 里的空记分：评审在准则要求的信息或动作缺失时仍给高分。
Paper analysisRead analysis
AttacksarXiv 2610.01367Oct 1
研究：高质量数据筛选挡不住投毒
提出 Bi-QSTO，优化能穿过质量筛选并在微调中削弱安全对齐的投毒样本
arXiv
2610.01367
Published
Oct 1
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama2-7B-Chat, Llama3-8B-Instruct, Qwen2-7B-Instruct and others5
- AttacksPoisoning and backdoors
- TechniqueGradient and representation optimization
- ComponentFine-tuning and open weights
Abstract质量筛选挡不住部分高分样本的安全削弱。
Bi-QSTO 研究一个实际训练缺口：基于质量的数据选择会先于微调决定样本去留，而以往微调投毒多假设毒样本直接进入训练集。已对齐 LLM 的拒答会被少量有害或看似无害的微调样本削弱。作者用三个问题串起实验：选择能否排除毒样本。
Paper analysisRead analysis
AttacksarXiv 2610.01365Oct 1
ReGap：无需真实隐私数据即可通过微调重新提取模型隐私关联
提出 ReGap，用自生成伪监督微调恢复模型中的隐私关联
arXiv
2610.01365
Published
Oct 1
Layer
Training and data layer
Setting
Models and APIs
Attacker
White-box, Gray-box
Evaluated models
GPT-2 Small, GPT-2 Medium, GPT-2 Large and others6
- AttacksExtraction and theft
- ComponentFine-tuning and open weights
Abstract提出无真实私有监督的恢复攻击 ReGap，实验显示微调可通过自生成数据唤醒模型潜在隐私关联，突显后适应隐私风险。
Paper analysisRead analysis
DefensesarXiv 2610.00685Oct 1
研究者提出用零空间投影净化 LoRA 微调 LLM 的后门
用零空间投影净化 LoRA 微调模型后门并保留基座与下游能力
arXiv
2610.00685
Published
Oct 1
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
LLaMA2-7B-Chat, LLaMA2-13B-Chat, Qwen2-7B-Instruct and others6
- DefensesModel hardening
- AttacksPoisoning and backdoors
- ComponentFine-tuning and open weights
Paper analysisRead analysis
DefensesarXiv 2609.39866Sep 30
研究者提出梯度封堵实现 LLM 发布前的预防性遗忘
提出 GSU 在发布前封堵敏感门梯度，抵抗下游微调获取禁止能力
arXiv
2609.39866
Published
Sep 30
Layer
Training and data layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
Llama-3.2-1B, Llama-3.2-3B, Llama-3.1-8B and others7
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
AbstractGSU 在发布前封堵 SiLU 门的获取梯度，并在 TOFU 与 WMDP 上降低下游微调得分。
Gradient-Sealed Unlearning（GSU）是一种发布前的 preemptive unlearning：不删除模型里已经有的能力，而是让指定禁止域能力更难被发布后的微调获取。
Paper analysisRead analysis
DefensesarXiv 2609.39279Sep 30
FDCU：用双重约束子空间投影抵御机器遗忘后的重训练攻击
提出 FDCU 双约束子空间投影遗忘，抵御再训练恢复有害知识
arXiv
2609.39279
Published
Sep 30
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Qwen2.5-3B, Llama-3-8B-Instruct, Qwen-3-8B
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
AbstractFDCU 用双掩码限制浅层抑制，在知识擦除与安全输出上提高再训练后的稳定性。
FDCU 是一种约束参数更新的遗忘方法，用来降低再训练把已抑制恶意行为重新引出的可能。
Paper analysisRead analysis

AI Safety Daily

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