Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 213papers have no tags yet and are not included in filtered results.

Impact：Harmful content Attack surface / entry point：Code repositories and dependencies Attack surface / entry point：Memory writes Attack surface / entry point：Training data Clear all

Another 213papers do not have category tags yet and are not included in filtered results.

DefensesarXiv 2610.13733Oct 9
KVeil: First-Place Solution for the ECCV 2026 LifeGenIP Unlearnable Video Challenge
Propose KVeil, which adds perturbations to face videos to prevent unauthorized customization from reproducing identity
arXiv
2610.13733
Published
Oct 9
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box, Black-box
Evaluated models
Wan2.2-TI2V-5B, HunyuanVideo-1.5, Kling 3.0 and others4
- AttacksMalicious use
- ComponentImage generation
AbstractKVeil adapts LETHE's reference-reading disruption to the challenge's two stages, ranking first on the leaderboard in both Final columns.
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
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 2610.01367Oct 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 2609.39788Sep 30
研究：多智能体系统潜在通信链路可被训练用于提升有害顺从
提出只训练或篡改多智能体潜在通信链路来提高有害遵从
arXiv
2609.39788
Published
Sep 30
Layer
Application layer
Setting
Multi-agent systems
Evaluated models
Llama-3.2-3B-Instruct, Qwen2.5-3B-Instruct, Gemma-3-1B-IT and others7
- AttacksModel tampering
- TechniqueGradient and representation optimization
Abstract潜在链路本身即可改变冻结智能体系统的安全，奖励既能攻击也能修复。
学习得到的潜在通信链路会改变由冻结、对齐智能体组成的系统是否遵从有害请求。。作者研究三种拓扑：双智能体、三级顺序协作，以及双专家加汇总。
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
DefensesarXiv 2609.36862Sep 29
VaccineBooster：面向有害微调对齐的混合扰动防御
提出 VaccineBooster，在对齐阶段混合嵌入扰动与梯度衰减抵御有害微调
arXiv
2609.36862
Published
Sep 29
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Llama-2-7B
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
AbstractVaccineBooster 把两种对齐阶段扰动合在一步里，内容更安全但拒答保留更少。
VaccineBooster 是一种对齐阶段防御：在提供方还看得到训练过程、还没接触用户微调数据时，把嵌入扰动和权重级梯度衰减合成同一次更新，用来应对之后不受信任的微调。有害微调可以只混入少量有害指令-回答对，就削弱拒答并提高有害输出的可能，任务效用却仍然好看。过滤会漏掉隐蔽样本，事后修复又往往不知道该修哪里。
Paper analysisRead analysis
DefensesarXiv 2610.00320Sep 29
研究显示少样本微调可绕过安全层局部冻结与修复防御
检验层冻结与奇异方向截断能否挡住少样本有害微调
arXiv
2610.00320
Published
Sep 29
Layer
Training and data layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
Llama-3.1-8B-Instruct, Llama-3.1-Tulu-3-8B-DPO, OLMo-2-1124-7B-Instruct and others10
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
Abstract定位与恢复可以复现，但知情攻击能把损伤移出冻结带，并打败跨检查点的截断修复。
这篇工作检验：把拒答定位到某些层或更新方向之后，能不能在攻击者知情时还守住拒答。。
Paper analysisRead analysis
OtherarXiv 2609.12413Sep 11
SoK 综述：智能体 AI 时代的越狱攻击、防御与实践考量
系统梳理智能体执行链路中的越狱攻击与防御并比较推理时防御
arXiv
2609.12413
Published
Sep 11
Layer
Application layer
Setting
AI Agent
Evaluated models
Qwen3.5-4B, Llama-3.1-8B-Instruct
- AttacksJailbreaks
- ComponentMemory
AbstractSoK 将越狱重放到智能体流水线，并用两种模型检验回复层防御的边际与代价。
Mia 等人把越狱安全从单轮回复扩展到智能体的规划、记忆、工具和智能体间通信，并问既有结论在更强原生对齐下是否还成立。
Paper analysisRead analysis
AttacksarXiv 2609.07048Sep 7
FreqDoor：面向视觉语言模型的频域后门攻击
提出 FreqDoor，在频域混合振幅并对生成式 VLM 做训练时视觉后门
arXiv
2609.07048
Published
Sep 7
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
BLIP-2 (opt-2.7b), InstructBLIP (flan-t5-xl), LLaVA (interleave-qwen)
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentVision
AbstractFreqDoor 用频域振幅混合给三款 VLM 植入仅视觉的训练时后门。
FreqDoor 是针对生成式 VLM 的仅视觉、训练时后门：在频域混合触发源振幅、保留干净图像相位，使触发在空间上分散。
Paper analysisRead analysis
InterpretabilityarXiv 2609.06934Sep 7
论文提出拒答的几何解释：事后安全训练脆弱，预训练期持续安全共训更抗攻击
解释后验安全的权重几何脆弱性，并提出预训练持续安全共训练
arXiv
2609.06934
Published
Sep 7
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen-2.5-7B, Qwen-2.5-7B-Instruct, Llama-3-8B and others15
- DefensesModel hardening
- AttacksModel tampering
- RiskInappropriate refusal
- ComponentFine-tuning and open weights
Abstract论文将后验安全脆弱性归因于正交抑制，并提出通过预训练持续共训练实现具备微调抗性的安全防护。
Paper analysisRead analysis
DefensesarXiv 2609.02293Sep 2
SEAL：通过共享专家对齐强化 MoE 全局安全
提出 SEAL，在 MoE 共享专家上做 DPO-LoRA 对齐以抵御越狱与权重篡改
arXiv
2609.02293
Published
Sep 2
Layer
Training and data layer
Setting
Models and APIs
Attacker
White-box, Black-box
Evaluated models
Qwen1.5-MoE-A2.7B, DeepSeekMoE-16B, GLM-4.7-Flash and others4
- DefensesModel hardening
- AttacksModel tampering
AbstractSEAL 把 DPO-LoRA 限制在混合 MoE 共享专家上，SEAL++再保护已有安全神经元方向。
SEAL是面向混合 MoE 的训练时防御：只对齐始终执行的共享专家，使安全不依赖路由器是否把 token 送到安全专家。
Paper analysisRead analysis
DefensesarXiv 2608.24354Aug 25
RACER：面向多模态大模型后门的区域感知一致性修复框架
提出 RACER，用区域感知一致性对抗微调去除多模态大模型后门
arXiv
2608.24354
Published
Aug 25
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
LLaVA-1.5-7B, Qwen2-VL-7B, Qwen2.5-VL-7B
- DefensesModel hardening
- AttacksPoisoning and backdoors
- ComponentVision
Paper analysisRead analysis
DefensesarXiv 2608.05045Aug 6
Gradient Immunity 提出 USG：在开放权重发布时阻断恶意微调梯度
提出 USG，在部分保护开源权重发布时阻断有害微调梯度
arXiv
2608.05045
Published
Aug 6
Layer
Training and data layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
Qwen3-14B（表中作 Qwen-14B）, Llama-3.1-8B（表中作 Llama-8B）
- DefensesModel hardening
- AttacksModel tampering
- ComponentFine-tuning and open weights
AbstractUSG 在 PPOW 下用零空间立方层阻断有害微调梯度。
Unidirectional Safety Gate 是面向 PPOW 发布设定的发布时防护：大部分权重仍可训练，一小部分安全关键组件由提供者侧加固保留，下游不必配合额外安全流程。。
Paper analysisRead analysis
AttacksarXiv 2608.01759Aug 3
EvoBreak 利用自我演化 Agent 的无害经验组合实施攻击
提出 EvoBreak，以良性任务组合诱导自进化 Agent 的持久经验并削弱安全边界
arXiv
2608.01759
Published
Aug 3
Layer
Application layer
Setting
AI Agent
Evaluated models
GPT-5-mini, Llama-3.1-8B-Instruct
- AttacksPoisoning and backdoors
- TechniqueAutomated search
- ComponentMemory
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2605.14605May 14
Study: malicious fine-tuning defenses fail under continued training, harmfulness rises across all 72 defense trajectories
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
Analysis and theoryarXiv 2604.02574Apr 3
Understanding the impact of safety unalignment on large language models
Comparing jailbreak fine-tuning and weight orthogonalization unalignment on harmful assistance and usefulness
arXiv
2604.02574
Published
Apr 3
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-4B-Instruct-2507, Llama-3.1-8B-Instruct, Qwen2.5-14B and others6
- AttacksModel tampering
- ComponentFine-tuning and open weights
Paper analysisRead analysis
AttacksarXiv 2603.08104Mar 9
Invisible Security Threat: Malicious Fine-Tuning of LLMs via Steganography
提出隐写恶意微调，使模型界面显良性并隐蔽传出有害内容
arXiv
2603.08104
Published
Mar 9
Layer
Training and data layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
GPT-4.1-2025-04-14, Llama-3.3-70B-Instruct, Phi-4 and others7
- AttacksModel tampering
- TechniqueEncoding and obfuscation
- ComponentFine-tuning and open weights
Paper analysisRead analysis

AI Safety Daily

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