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

Oct 12 · Monday

All papers

Found 164 papers

Filter4

System component

MCP and skills12 Memory39 RAG3 Guardrails and judges13 Monitors5 Reasoning chains9 Permissions and sandboxing19 Fine-tuning and open weights37 Vision7 Video3 Multimodal13

Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 247papers have no tags yet and are not included in filtered results.

Attack surface / entry point：Code repositories and dependencies Attack surface / entry point：Memory writes Attack surface / entry point：Inter-agent messages Attack surface / entry point：Training data Clear all

Another 247papers 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
Analysis and theoryarXiv 2610.13051Oct 9
Space beats statistics, retraining beats ranking: auditing training labels under label-only backdoors
Auditing training labels for label-only backdoors, comparing spaces and statistics, and judging by retraining
arXiv
2610.13051
Published
Oct 9
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
ResNet-32
- DefensesDetection and filtering
- AttacksPoisoning and backdoors
AbstractLabel inspection must be done in a space that has not been fitted to these labels, and judged by retrained models rather than detection scores.
The authors ask: can a defender without trusted data decide which training labels to keep, and how much clean data this destroys, and what the final model accuracy is.
Paper analysisRead analysis
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 2610.10345Oct 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 2610.10000Oct 7
超越奖励压制：有界奖励热启动 Bandit 的近最优离线攻击
提出热启动 Bandit 的部署前有界离线注入攻击
arXiv
2610.10000
Published
Oct 7
Layer
Training and data layer
Setting
LLM applications
Evaluated models
UCB, Thompson Sampling, ϵ-greedy
- AttacksPoisoning and backdoors
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 2610.09264Oct 7
研究者提出 PackHallu：经规则文件提示注入对编码 Agent 发起包幻觉攻击
提出 PackHallu，经规则文件提示注入诱导编码 Agent 换用攻击者指定包
arXiv
2610.09264
Published
Oct 7
Layer
Application layer
Setting
coding agent
Evaluated models
Qwen2.5-Coder-7B, Qwen3-Coder-30B, Devstral-Small-2-24B and others19
- AttacksPrompt injection
- TechniqueAutomated search
AbstractPackHallu 诱导规则文件换包。
PackHallu 研究一种针对编程智能体规则文件的提示注入：让智能体在本应使用合法依赖时，改导入攻击者指定并控制的包。
Paper analysisRead analysis
DefensesarXiv 2610.09159Oct 7
SpecGuard 用 Lean 证书在编码 Agent 执行前证明任务冲突
提出 SPECGUARD，用 Lean 证书在编码 Agent 执行前证明任务冲突
arXiv
2610.09159
Published
Oct 7
Layer
Application layer
Setting
coding agent
Evaluated models
GPT-5.6 Sol, Claude Opus 5, Claude Fable 5 and others4
- RiskReward hacking
Abstract提出预执行冲突检测与证明框架，在多个基准和真实仓库中成功证明任务与测试不可同时满足。
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
DefensesarXiv 2610.07657Oct 6
DEFER1 多智能体防御研究报告受控测试中的攻击成功率下降
实现 DEFER，以确定性检查优先拦截多智能体流水线的内容攻击
arXiv
2610.07657
Published
Oct 6
Layer
Application layer
Setting
Multi-agent systems
Evaluated models
Qwen3-235B-A22B-Instruct-2507, gpt-oss-120b, Llama-4-Scout-17B-16E-Instruct and others5
- DefensesPermissions and approval
- AttacksPrompt injection
- ComponentGuardrails and judges
AbstractDEFER 把确定性检查放在 judge 之前。
DEFER 用来测量多智能体授权流水线里，确定性规则在何处结束、LLM judge 从何处开始。
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
AttacksarXiv 2610.07645Oct 6
SkillPoison：用成功经验投毒自进化 Agent 的技能形成
提出 SkillPoison，用局部有效的成功经验诱导自进化 Agent 形成过度泛化技能
arXiv
2610.07645
Published
Oct 6
Layer
Application layer
Setting
AI Agent
Evaluated models
deepseek-v4-flash, Qwen V3.8 Flash, GPT-5.5
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentMemory
AbstractSkillPoison 通过局部有效且验证成功的经验诱导技能过度泛化，在多个评测基准上实现了持久的下游攻击影响。
SkillPoison 提出了一种针对自进化智能体经验到技能形成过程的攻击范式，通过组织已验证的成功经验诱导技能过度泛化。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2610.07639Oct 6
HarnessSecurity-Bench horizontally evaluates the security mechanisms of six coding agent frameworks.
评测编码智能体 harness 安全机制在对抗任务中的防护效果
arXiv
2610.07639
Published
Oct 6
Layer
Application layer
Setting
coding agent
- AttacksPrompt injection
- ComponentPermissions and sandboxing
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 2610.07510Oct 6
PersistBD: attackers can strengthen backdoors before release so that they maintain high attack success rates even after developer SFT and RL.
提出 PERSISTBD，发布前强化供应链后门使其经良性 SFT 与 RL 后仍外发凭据
arXiv
2610.07510
Published
Oct 6
Layer
Application layer
Setting
coding agent
Evaluated models
Qwen2.5-Coder-3B-Instruct, Qwen2.5-Coder-7B-Instruct, Qwen3-Coder-30B-A3B-Instruct
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
Abstract总结了论文关于后门在 SFT/RL 中存活的发现、两因子机制、PERSISTBD 方法与核心数字。
本文研究第三方 LLM 智能体在开发者进行良性监督微调（SFT）和强化学习（RL）后训练时的后门持久性风险。
Paper analysisRead analysis
DefensesarXiv 2610.06636Oct 6
差分隐私混合公共数据集提升隐私学习效果
提出 DP-MixMin，私有学习公开数据混合并做差分隐私微调
arXiv
2610.06636
Published
Oct 6
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
ResNet-18, GPT-2 Medium
- AttacksExtraction and theft
AbstractDP-MixMin 私下学公开混合再 DP 微调，胸片 macro AUC 与邮件困惑度优于均匀混合。
DP-MixMin 是一条端到端的差分隐私训练流程：先私下决定多个公开数据集的预训练混合，再预训练并在敏感数据上 DP 微调。作者称这是首个这样的流程。
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
DefensesarXiv 2610.05640Oct 5
Research reveals that the CaMeL defense fails in multi-agent systems and proposes multi-CaMeL
提出 multi-CaMeL，在智能体调用边界隔离可信指令与非可信数据
arXiv
2610.05640
Published
Oct 5
Layer
Application layer
Setting
Multi-agent systems
Evaluated models
Claude Fable 5, GPT-5.5, GPT-4.1 and others6
- DefensesInstruction–data separation
- AttacksPrompt injection
Abstract论文针对多智能体中单体防御不组合的边界清洗问题，提出 multi-CaMeL 协议彻底阻断了 MultiAgentDojo 上的注入攻击。
Paper analysisRead analysis
InterpretabilityarXiv 2610.05295Oct 4
研究用因果追踪区分间接提示注入中的可读信号与有效干预
用因果追踪区分间接提示注入中的角色可读性与有效干预
arXiv
2610.05295
Published
Oct 4
Layer
Application layer
Setting
AI Agent
Evaluated models
Qwen-2.5-7B, Qwen-2.5-1.5B, Llama-3.1-8B and others8
- AttacksPrompt injection
Abstract作者称源角色先于强工具选择效应可读，有效编辑还取决于位置、长度和方向来源。
在权重固定的白盒模型上，源角色可以先被读出。沿相关方向的编辑是否改变行为，要在目标位点上另测。。
Paper analysisRead analysis
AttacksarXiv 2610.05139Oct 4
研究揭示代码注释中的上下文注入攻击面
提出推理时上下文注入，用代码注释诱导补全生成含弱点的代码
arXiv
2610.05139
Published
Oct 4
Layer
Prompt layer
Setting
coding agent
Evaluated models
StarCoderBase, DeepSeek-Coder-base, Qwen2.5-base and others10
- AttacksPrompt injection
Paper analysisRead analysis

AI Safety Daily

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

All papers

Attack type

Attack technique

Model risks

Defense type

Impact

System type

Attacker access

Attack surface / entry point

Layer

System component

KVeil: First-Place Solution for the ECCV 2026 LifeGenIP Unlearnable Video Challenge

Space beats statistics, retraining beats ranking: auditing training labels under label-only backdoors

Cancerous backdoor attack: progressive activation via invisible tokens and a stage controller.

SLDR：通过选择性层恢复与动态路由防御恶意微调

超越奖励压制：有界奖励热启动 Bandit 的近最优离线攻击

OrthoPurify 用一步正交投影净化被后门污染的视觉语言模型

研究者提出 PackHallu：经规则文件提示注入对编码 Agent 发起包幻觉攻击

SpecGuard 用 Lean 证书在编码 Agent 执行前证明任务冲突

Researchers propose answer-side backdoors: letting the model self-generate trigger words to bypass safety refusals.

DEFER1 多智能体防御研究报告受控测试中的攻击成功率下降

研究揭示策略蒸馏的安全后门风险

SkillPoison：用成功经验投毒自进化 Agent 的技能形成

HarnessSecurity-Bench horizontally evaluates the security mechanisms of six coding agent frameworks.

TRACE：仅凭 checkpoint 更新审计大语言模型有害目标

PersistBD: attackers can strengthen backdoors before release so that they maintain high attack success rates even after developer SFT and RL.

差分隐私混合公共数据集提升隐私学习效果

研究者训练出按多智能体拓扑触发的代码后门模型

Research reveals that the CaMeL defense fails in multi-agent systems and proposes multi-CaMeL

研究用因果追踪区分间接提示注入中的可读信号与有效干预

研究揭示代码注释中的上下文注入攻击面