AttacksarXiv 2610.09973Oct 7
从期望危害到似然：LLM 智能体越狱的概率重构
提出 OPUR 攻击，用似然梯度优化对抗后缀越狱 LLM 智能体
arXiv
2610.09973
Published
Oct 7
Layer
Application layer
Setting
AI Agent
Attacker
White-box
Evaluated models
Llama-3.1-8B-Instruct, Ministral-8B-Instruct-2410
- AttacksJailbreaks
- TechniqueGradient and representation optimization
Paper analysisRead analysis
AttacksarXiv 2610.06670Oct 6
研究者提出 Reward Stealing Attack，从对齐模型行为反推安全奖励实施攻击
提出 ReSA，从对齐模型行为恢复安全奖励并在解码时反转
arXiv
2610.06670
Published
Oct 6
Layer
Model layer
Setting
Models and APIs
Attacker
White-box, Gray-box
Evaluated models
Llama3.1-8B-Instruct, Gemma-7B-Instruct, Qwen2.5-7B-Instruct and others9
- AttacksJailbreaks
- TechniqueDefense-targeted
AbstractReSA 用最大熵 IRL 从对齐模型行为恢复代理安全奖励，并在解码时反转以诱导有害输出。
ReSA 是一种对抗攻击：从对齐模型的可观察行为恢复代理安全奖励，再在解码时把该奖励反过来用。。
Paper analysisRead analysis
AttacksarXiv 2610.05943Oct 5
CRIME 框架利用良性 Agent 技能组合诱发恶意行为
提出 CRIME 框架，组合良性 Agent 技能诱发恶意行为
arXiv
2610.05943
Published
Oct 5
Layer
Application layer
Setting
AI Agent
Evaluated models
DeepSeek-V4-Flash, GPT-4o-mini, GPT-4o and others6
- AttacksJailbreaks
- TechniqueAutomated search
- ComponentMCP and skills
Abstract良性技能经 DCA 或 ACA 组合后可产生恶意环境后果，单独审查对不上。
CRIME 把公共技能库中各自通过审查的技能配成两两组合，看联合执行能否实现指定恶意任务。。
Paper analysisRead analysis
AttacksarXiv 2609.15383Sep 14
微软团队提出 capability laundering 攻击：弱模型拆分任务向对齐前沿模型套取能力
提出能力洗钱攻击，拆分任务向对齐模型套取能力
arXiv
2609.15383
Published
Sep 14
Layer
Application layer
Setting
AI Agent
Attacker
Black-box
Evaluated models
GPT-5.5, Claude Opus 4.8, Grok-4.3 and others7
- AttacksJailbreaks
- TechniqueMulti-turn escalation
Abstract作者提出了能力洗钱攻击，发现未对齐编排器分解咨询对齐模型可恢复能力差距，指出需建立跨会话防御。
Paper analysisRead analysis
AttacksarXiv 2609.02172Sep 2
BOSS：以广度优先后缀搜索改进 GCG 越狱优化
提出广度优先后缀搜索 BOSS，改进 GCG 越狱优化
arXiv
2609.02172
Published
Sep 2
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
Llama-2-7B-Chat, Yi-1.5-9B-Chat, Qwen2-7B-Instruct and others8
- AttacksJailbreaks
- TechniqueGradient and representation optimization
AbstractBOSS 把 GCG 类优化从单条深搜索改为多条短轨迹加源侧诊断。
BOSS 是插在 GCG 类离散 suffix 优化上的搜索框架：局部梯度更新不变，变的是固定预算花在深度还是广度。
Paper analysisRead analysis
AttacksarXiv 2609.05525Sep 2
DIVA：针对离散扩散视觉语言模型的跨步条件传播视觉越狱
提出 DIVA，利用跨步条件传播对离散扩散视觉语言模型实施视觉越狱
arXiv
2609.05525
Published
Sep 2
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
LLaDA-V, MMaDA, LLaDA2.0-Uni
- AttacksJailbreaks
- TechniqueGradient and representation optimization
- ComponentVision
Paper analysisRead analysis
AttacksarXiv 2608.07430Aug 8
研究者提出 SN-Guided Diffusion，利用扩散大模型安全神经元实现离线越狱
提出 SN-Guided Diffusion，利用安全神经元实现离线越狱
arXiv
2608.07430
Published
Aug 8
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box, White-box
Evaluated models
Qwen2.5-7B-Instruct, LLaDA-8B-Instruct, Dream-Instruct-7B and others15
- AttacksJailbreaks
- TechniqueGradient and representation optimization
Abstract论文揭示了扩散语言模型对自回归安全神经元的继承性，并提出纯离线黑盒越狱框架 SN-Guided Diffusion。
论文系统评估了扩散大语言模型（DLLM）在安全对齐中的双重角色——既作为白盒攻击目标，又作为黑盒越狱生成的对抗工具。
Paper analysisRead analysis
AttacksarXiv 2607.23496Jul 26
Concept2Scenario：用 SAE 概念归因发现可迁移的越狱场景
提出 Concept2Scenario，借助内部概念归因发现可迁移的越狱场景
arXiv
2607.23496
Published
Jul 26
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box, Black-box
Evaluated models
Qwen3.5-9B, Llama-3.1-8B, Ministral-3-8B and others6
- AttacksJailbreaks
- TechniqueGradient and representation optimization
Abstract该研究从表征空间因果归因揭示场景削弱拒绝的机制，发现可复用脆弱场景先验，提升多模型黑盒越狱效能与探索速度。
Paper analysisRead analysis
AttacksarXiv 2606.21077Jun 19
OTTER：仅替换五个 token 即可绕过 GPT 系列毒性审核
提出 OTTER 词元替换越狱，绕过表面毒性审核诱导有害回答
arXiv
2606.21077
Published
Jun 19
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
gpt-4-turbo, gpt-4o, gpt-4o-mini and others4
- AttacksJailbreaks
- TechniqueEncoding and obfuscation
- ComponentGuardrails and judges
Abstract论文展示了表面毒性与恶意意图的解耦现象，通过标准 API 实现黑盒越狱测试，并提出闭环分类器加固建议。
Paper analysisRead analysis
AttacksarXiv 2606.09084Jun 8
研究者提出 Context-Fractured Decomposition 攻击
提出 CFD 跨会话分解攻击，利用工件来源缺口绕过工具型 Agent 的本地检查
arXiv
2606.09084
Published
Jun 8
Layer
Application layer
Setting
AI Agent
Evaluated models
GPT-4.1, GPT-5-nano, Claude-3.5-Sonnet and others6
- AttacksJailbreaks
- TechniqueDormant triggers
- ComponentMemory
AbstractCFD 用无指令 artifact 把越狱拆到跨会话。
CFD针对工具型 LLM 智能体：利用 artifact 来源不被跟踪这一部署失败模式，把有害目标拆开后延迟组合。
Paper analysisRead analysis
AttacksarXiv 2606.05609Jun 4
SlotGCG 利用提示词位置漏洞提升 LLM 越狱攻击成功率
提出 SlotGCG，按注意力选槽插入对抗 token 以越狱指令模型
arXiv
2606.05609
Published
Jun 4
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
LLAMA2-7B-CHAT, LLAMA2-13B-CHAT, LLAMA-3.1-8B-INSTRUCT and others10
- AttacksJailbreaks
- TechniqueGradient and representation optimization
Paper analysisRead analysis
AttacksarXiv 2605.20519May 20
CodecAttack：在音频编解码器潜空间构造对抗扰动，绕过有损压缩防御
提出 CodecAttack，在编解码器隐空间构造可穿透有损压缩的对抗音频
arXiv
2605.20519
Published
May 20
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
Qwen2-Audio-7B-Instruct, Audio Flamingo 3, Qwen2.5-Omni
- AttacksJailbreaks
- TechniqueGradient and representation optimization
- ComponentAudio
Paper analysisRead analysis
AttacksarXiv 2605.07399May 8
研究者提出 GPO-V，用全局概率优化越狱扩散视觉语言模型
提出 GPO-V，用全局概率优化的视觉扰动越狱扩散视觉语言模型
arXiv
2605.07399
Published
May 8
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box, Black-box
Evaluated models
LLaDA-V, LaViDA
- AttacksJailbreaks
- TechniqueGradient and representation optimization
- ComponentVision
Abstract作者用视觉扰动改 dVLM 的全局生成概率，以绕过两类拒绝。
GPO-V 把越狱从“开头必须顺从”改成“整段去噪概率朝肯定方向走”。。
Paper analysisRead analysis
AttacksarXiv 2605.04446May 6
Misrouter：利用 MoE 路由机制对混合专家大模型发起仅输入攻击
提出 Misrouter，借 MoE 路由优化仅输入前缀并迁移攻击
arXiv
2605.04446
Published
May 6
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box, White-box
Evaluated models
GPT-4o mini, Qwen-Plus, Phi-4 and others18
- AttacksJailbreaks
- TechniqueGradient and representation optimization
Paper analysisRead analysis
AttacksarXiv 2604.24983Apr 28
PEO：直接优化原始提示词嵌入的 LLM 白盒越狱方法
提出 PEO，直接优化原提示词嵌入实现白盒越狱
arXiv
2604.24983
Published
Apr 28
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box
Evaluated models
Vicuna-7B-v1.3, Llama-2-7B-Chat, Qwen3-4B-Instruct and others4
- AttacksJailbreaks
- TechniqueGradient and representation optimization
AbstractPEO 原位优化提示词嵌入、多轮只重攻失败样本。
PEO是一种不追加对抗 token 的 white-box 越狱：直接改已有提示词嵌入，并只对未成功样本更换续写目标再优化。
Paper analysisRead analysis
AttacksarXiv 2603.13847Mar 14
SWhisper 用近超声隐蔽声学通道对语音驱动 LLM 实现黑盒越狱
提出 SWhisper，用近超声波向语音驱动 LLM 隐蔽注入越狱提示
arXiv
2603.13847
Published
Mar 14
Layer
Application layer
Setting
LLM applications
Attacker
Black-box
Evaluated models
Meta-Llama-3.1-8B-Instruct, Gemma-3-4B, Qwen2.5-7B-Instruct and others9
- AttacksJailbreaks
- TechniqueCross-modal payloads
- ComponentAudio
Paper analysisRead analysis
AttacksarXiv 2603.01414Mar 2
通过动作级操纵越狱具身 LLM
提出 Blindfold，用动作级操纵越狱具身 LLM
arXiv
2603.01414
Published
Mar 2
Layer
Application layer
Setting
Embodied systems and robotics
Attacker
No-box (offline transfer)
Evaluated models
Llama-3.1-8B, DeepSeek-R1-14B, Gemma-3-27B and others8
- AttacksJailbreaks
- TechniqueEncoding and obfuscation
AbstractBlindfold 以动作级代理规划越狱具身 LLM，作者称可绕过语义检查并在仿真与真机生效。
Blindfold 是面向具身 LLM 规划模块的 no-box 攻击，把恶意意图改写成看似良性的动作序列。
Paper analysisRead analysis
AttacksarXiv 2602.10179Feb 11
研究者提出视觉中心越狱攻击 VJA
提出纯视觉越狱攻击 VJA，把有害编辑指令编码进图像
arXiv
2602.10179
Published
Feb 11
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
Nano Banana Pro, GPT Image 1.5, Qwen-Image-Edit and others9
- AttacksJailbreaks
- TechniqueCross-modal payloads
- ComponentImage generation
Paper analysisRead analysis
AttacksarXiv 2602.07107Feb 7
ShallowJail: Guiding Jailbreak Attacks Against Large Language Models
提出 ShallowJail，在推理时用转向向量改写隐藏状态实现越狱
arXiv
2602.07107
Published
Feb 7
Layer
Model layer
Setting
Models and APIs
Evaluated models
Qwen3-4B-Instruct-2507, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct
- AttacksJailbreaks
- TechniqueGradient and representation optimization
Paper analysisRead analysis
AttacksarXiv 2602.03265Feb 3
Beyond Suffixes: The Impact of Token Position in GCG Adversarial Attacks on Large Language Models
比较 GCG 对抗 token 作为前缀与后缀的越狱效果
arXiv
2602.03265
Published
Feb 3
Layer
Prompt layer
Setting
Models and APIs
Attacker
White-box, Black-box
Evaluated models
deepseek-llm-7b-chat, Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3 and others5
- AttacksJailbreaks
- TechniqueGradient and representation optimization
Paper analysisRead analysis

AI Safety Daily

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