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

Oct 13 · Tuesday

All papers

Found 55 papers

Filter6

System type

Models and APIs31 LLM applications4 AI Agent20 coding agent5

Attacker access

Black-box23

Attack surface / entry point

User input42 Tool responses5 Tool and skill descriptions12 Retrieval corpus3 Images, audio and video5 Distribution channels9

Layer

Prompt layer30 Model layer1 Application layer24

System component

MCP and skills11 Guardrails and judges7 Reasoning chains4 Permissions and sandboxing3 Vision5 Multimodal6

Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 201papers have no tags yet and are not included in filtered results.

Attack technique：Multi-turn escalation Attack technique：Conversation structure manipulation Attack technique：Selection manipulation Layer：Prompt layer Layer：Model layer Layer：Application layer Clear all

Another 201papers do not have category tags yet and are not included in filtered results.

AttacksarXiv 2610.12524Oct 8
An exploration of attacks on AI agent systems
Retest and extend prompt injection against AI agents, and use Agent Card deception to manipulate A2A selection
arXiv
2610.12524
Published
Oct 8
Layer
Application layer
Setting
AI Agent
Evaluated models
Ollama Mistral 0.9.5, llama3-8b-8192, gemini-2.0-flash-lite and others5
- AttacksPrompt injection
- TechniqueEncoding and obfuscation
- ComponentMCP and skills
AbstractSingle-agent filtering and classification can be bypassed by format or wording, and A2A card selection is also biased toward names and descriptions.
The authors retested prompt injection on their own LangChain, AutoGen, and A2A systems, and modified the Agent Card to see if choices shifted. The abstract states the experiments mainly highlight single-agent weaknesses while extending attacks to multi-agent.
Paper analysisRead analysis
AttacksarXiv 2610.08098Oct 6
CORSA：面向技能路由器的技能注入攻击可跨路由器与 LLM 迁移
提出 CORSA，按任务簇优化技能注入的检索与执行
arXiv
2610.08098
Published
Oct 6
Layer
Application layer
Setting
coding agent
Evaluated models
GPT-5.4, DeepSeek-V4-Pro, Qwen3.8-27B and others5
- AttacksPoisoning and backdoors
- TechniqueSelection manipulation
- ComponentMCP and skills
Paper analysisRead analysis
AttacksarXiv 2610.04985Oct 4
研究揭示 Jev 作为应用决策层的安全与隐私风险
评估应用决策层的输入注入与隐私推断风险
arXiv
2610.04985
Published
Oct 4
Layer
Application layer
Setting
LLM applications
Attacker
Black-box, Gray-box
Evaluated models
Jev, NanoJev
- AttacksPrompt injection
- TechniqueRole-play and persuasion
AbstractJev 的类型化决策可被输入与训练操纵，也会泄露信息，并同时支持检测与滥用。
Jev 是 TypeSafe 的 System One 模型，把应用 state 和固定问题变成 Choice、Noul 或 Score 的类型化决策与概率。作者用官方 API 和可改训练的 NanoJev，考察它作为决策层时的操纵、泄露与双重用途。
Paper analysisRead analysis
AttacksarXiv 2610.00392Sep 30
A2A-TIBA：针对 A2A 协议 LLM 智能体的间接提示注入攻击与三层同构攻防模型
提出 A2A-TIBA，经 A2A 对等任务间接注入并植入常驻回调
arXiv
2610.00392
Published
Sep 30
Layer
Application layer
Setting
Multi-agent systems
Evaluated models
MiniMax-M3, deepseek-v4-pro, glm-5.3 and others8
- AttacksPrompt injection
- TechniqueMulti-turn escalation
- ComponentMCP and skills
AbstractA2A-TIBA 一次植入后绕过智能体，GDA 用四类结果定位防线，ELA-ITL 把信封标注列为第三层。
A2A-TIBA 把间接提示注入和绕过放进 A2A：gray-box 攻击者已能投递任务，诱导目标部署常驻回调程序后，下令和外传不再经过智能体。
Paper analysisRead analysis
AttacksarXiv 2609.36956Sep 29
The researchers propose Controlled Decoding Attacks, jailbreaking black-box LLMs through a text-only interface
提出 BLINDBIAS，通过纯文本续写与前缀控制越狱黑盒模型
arXiv
2609.36956
Published
Sep 29
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
GLM-5, Gemini-3.5-Flash, Qwen3-32B and others4
- AttacksJailbreaks
- TechniqueConversation structure manipulation
Abstract作者认为，隐藏数值概率并不足以消除支持重复采样和前缀续写的解码攻击面。
BLINDBIAS 是一种针对 sample-only 续写接口的 black-box 解码时越狱框架，研究没有目标数值概率时，采样能否支持逐步控制。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.36849Sep 29
Study evaluates GradSafe's jailbreak detection vulnerability in multi-turn dialogues
评测多轮对话中 GradSafe 越狱检测的可分性与脆弱性
arXiv
2609.36849
Published
Sep 29
Layer
Prompt layer
Setting
Models and APIs
Evaluated models
Llama 3.1 8B Instruct, Qwen2.5-7B-Instruct
- DefensesDetection and filtering
- AttacksJailbreaks
- TechniqueMulti-turn escalation
- ComponentGuardrails and judges
AbstractCWS 显示不安全梯度在真实多轮流量中变弱，且随攻击族和模型而变。
作者评估梯度越狱检测从单条提示词迁到多轮对话后还剩多少可分性，方法是用 Context Window Scanner 把 GradSafe 原封不动地套到用户轮窗口上。
Paper analysisRead analysis
AttacksarXiv 2609.30383Sep 25
研究者提出技能级联攻击
提出技能级联攻击，把有害目标拆散到多个技能里绕过单技能扫描
arXiv
2609.30383
Published
Sep 25
Layer
Application layer
Setting
AI Agent
Evaluated models
GPT-5.4, GPT-5.4 Mini, Claude Opus 4.6 and others11
- AttacksPoisoning and backdoors
- TechniqueDefense-targeted
- ComponentMCP and skills
Abstract论文探讨了多技能智能体系统面临的技能级联攻击风险，提出自动化红队框架与基准集，指出跨技能交互存在安全防护盲区。
Paper analysisRead analysis
AttacksarXiv 2609.29775Sep 24
输出前缀攻击瞄准推理模型
提出推理通道预填与输出前缀组合攻击以越狱推理模型
arXiv
2609.29775
Published
Sep 24
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
Gemini 3 Flash Preview, DeepSeek V4 Flash, Claude Haiku 4.5
- AttacksJailbreaks
- TechniqueReasoning-chain manipulation
- ComponentReasoning chains
Paper analysisRead analysis
AttacksarXiv 2609.27542Sep 23
Researchers use control token injection to delete gpt-oss-20b's reasoning chain and bypass chain-of-thought monitoring
提出控制标记注入，清空工具 Agent 的思维链以绕过推理监控
arXiv
2609.27542
Published
Sep 23
Layer
Application layer
Setting
AI Agent
Evaluated models
openai/gpt-oss-20b, google/gemma-4-26B-A4B-it, deepseek-ai/DeepSeek-R1-Distill-Qwen-7B and others4
- AttacksPrompt injection
- TechniqueConversation structure manipulation
- ComponentReasoning chains
Abstract论文阐明控制标记注入可抑制 CoT 使监控失效并促成拒绝绕过，且工具执行受解析器宽容度影响。
Paper analysisRead analysis
AttacksarXiv 2609.26761Sep 23
A2M: A two-stage agent hijacking framework targeting the MCP ecosystem
提出两阶段框架 A2M，优化 MCP 工具元数据与返回载荷以劫持 Agent
arXiv
2609.26761
Published
Sep 23
Layer
Application layer
Setting
AI Agent
Evaluated models
GLM-4.6, Qwen3-Max, DeepSeek-V3.1 and others5
- AttacksPrompt injection
- TechniqueAutomated search
- ComponentMCP and skills
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 2609.08373Sep 8
结构越狱可跨厂商泛化但不叠加：IICL 的跨厂商与多语言研究
检验结构越狱 IICL 能否跨厂商泛化并与语言约束叠加
arXiv
2609.08373
Published
Sep 8
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
gemini-2.5-flash, gemini-2.5-flash-lite
- AttacksJailbreaks
- TechniqueConversation structure manipulation
Paper analysisRead analysis
AttacksarXiv 2609.03884Sep 3
HookPry：利用插件生命周期钩子更新劫持 AI Agent 框架
提出 HOOKPRY，用插件钩子版本更新供应链劫持 Agent 框架
arXiv
2609.03884
Published
Sep 3
Layer
Application layer
Setting
AI Agent
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentMCP and skills
Paper analysisRead analysis
AttacksarXiv 2609.03247Sep 3
研究揭示大语言模型的自发身份认证失败：五款模型的分阶段开发者身份实验
提出 MIPC 与 CFA，观察模型自拟测验后接受开发者身份
arXiv
2609.03247
Published
Sep 3
Layer
Prompt layer
Setting
Models and APIs
Evaluated models
ChatGPT, Claude, Qwen3.7-Plus and others5
- AttacksJailbreaks
- TechniqueMulti-turn escalation
Abstract自拟测验可造成对话层虚假认证，但所测授权边界并未因此改变。
这项工作把“模型自己出题、自己批改、再宣布身份已核验”定义成对话层认证失败，并与后端提权分开。
Paper analysisRead analysis
AttacksarXiv 2609.02414Sep 2
Blueprint 框架用世界观模拟与 18 个影响因子提升多轮越狱成功率
提出多轮越狱框架 BLUEPRINT，用世界观模拟与心理因素诱导有害输出
arXiv
2609.02414
Published
Sep 2
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
Qwen3-Next-80B, DeepSeek-V3.2, GLM-4.7 and others6
- AttacksJailbreaks
- TechniqueMulti-turn escalation
AbstractBLUEPRINT 解耦多轮越狱策略与情境模拟，揭示任务具象化是摆脱拒绝的核心机制，为多轮安全防御提供新视角。
BLUEPRINT 是一个将因子化社会影响力策略空间与跨轮世界观模拟解耦的多轮大语言模型安全评测框架。
Paper analysisRead analysis
AttacksarXiv 2609.02035Sep 2
ISM：通过语义匹配隐式操纵 LLM Agent 的技能选择
提出 ISM，用语义匹配隐式操纵 Agent 技能选择
arXiv
2609.02035
Published
Sep 2
Layer
Application layer
Setting
AI Agent
Evaluated models
Llama-2-7B, Llama-3-8B, Llama-3.1-8B and others8
- AttacksPoisoning and backdoors
- TechniqueSelection manipulation
- ComponentMCP and skills
Paper analysisRead analysis
AttacksarXiv 2608.27800Aug 28
ContextLeak：通过恶意工具窃取 LLM Agent 运行时上下文
提出 ContextLeak，用强化学习优化恶意工具描述，诱导 Agent 外泄运行时上下文
arXiv
2608.27800
Published
Aug 28
Layer
Application layer
Setting
AI Agent
Evaluated models
Qwen-3-8B, GPT-OSS-20B, Gemma-4-E4B-it and others11
- AttacksPrompt injection
- TechniqueSelection manipulation
- ComponentMCP and skills
Paper analysisRead analysis
AttacksarXiv 2608.25776Aug 26
EvoMal attack uses self-poisoning of self-evolving coding agents to achieve worm-like propagation
提出 EVOMAL，诱导自演化编码 Agent 在创建技能时自我投毒并蠕虫传播
arXiv
2608.25776
Published
Aug 26
Layer
Application layer
Setting
coding agent
Evaluated models
Devstral-Small-2, Gemma-4-31B-IT, Qwen3-Coder-Next and others7
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentMCP and skills
Abstract论文指出了自演化智能体模仿检索代码的自我毒化漏洞，提出 EVOMAL 攻击与防御，阐明了供应链新风险。
Paper analysisRead analysis
AttacksarXiv 2608.12273Aug 13
研究者提出 Convergent Detour Hijacking 攻击，可让技能型 LLM Agent 绕路并放大资源消耗
提出 CDH 攻击，让技能型 Agent 绕路放大资源消耗
arXiv
2608.12273
Published
Aug 13
Layer
Application layer
Setting
AI Agent
Evaluated models
Claude-Haiku-4.5, DeepSeek-V4-Pro, DeepSeek-V4-Flash and others6
- AttacksPrompt injection
- TechniqueSelection manipulation
- ComponentMCP and skills
AbstractCDH 让静态技能在保持任务完成时拉高调用与 token，耗时则因后端而异。
CDH 是针对渐进披露技能智能体的发布者侧、纯文本攻击：一个静态协调器把本可正确完成的任务，拐进有界但更耗资源的技能绕路，再交还原任务。
Paper analysisRead analysis
AttacksarXiv 2608.09867Aug 11
Researchers steal proprietary LLM inference chains via cross-model compatibility of encrypted inference blocks
利用加密思维块的跨模型重放，诱导弱模型逐字转写专有推理链
arXiv
2608.09867
Published
Aug 11
Layer
Prompt layer
Setting
Models and APIs
Attacker
Black-box
Evaluated models
Claude Opus 4.8, Claude Sonnet 4.6, Claude Haiku 4.5 and others13
- AttacksExtraction and theft
- TechniqueReasoning-chain manipulation
- ComponentReasoning chains
Abstract论文揭示了客户端加密思维块跨模型互换漏洞，提出了低成本提取方法并给出防御方案。
Paper analysisRead analysis

AI Safety Daily

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

All papers

Attack type

Attack technique

Defense type

Impact

System type

Attacker access

Attack surface / entry point

Layer

System component

An exploration of attacks on AI agent systems

CORSA：面向技能路由器的技能注入攻击可跨路由器与 LLM 迁移

研究揭示 Jev 作为应用决策层的安全与隐私风险

A2A-TIBA：针对 A2A 协议 LLM 智能体的间接提示注入攻击与三层同构攻防模型

The researchers propose Controlled Decoding Attacks, jailbreaking black-box LLMs through a text-only interface

Study evaluates GradSafe's jailbreak detection vulnerability in multi-turn dialogues

研究者提出技能级联攻击

输出前缀攻击瞄准推理模型

Researchers use control token injection to delete gpt-oss-20b's reasoning chain and bypass chain-of-thought monitoring

A2M: A two-stage agent hijacking framework targeting the MCP ecosystem

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

结构越狱可跨厂商泛化但不叠加：IICL 的跨厂商与多语言研究

HookPry：利用插件生命周期钩子更新劫持 AI Agent 框架

研究揭示大语言模型的自发身份认证失败：五款模型的分阶段开发者身份实验

Blueprint 框架用世界观模拟与 18 个影响因子提升多轮越狱成功率

ISM：通过语义匹配隐式操纵 LLM Agent 的技能选择

ContextLeak：通过恶意工具窃取 LLM Agent 运行时上下文

EvoMal attack uses self-poisoning of self-evolving coding agents to achieve worm-like propagation

研究者提出 Convergent Detour Hijacking 攻击，可让技能型 LLM Agent 绕路并放大资源消耗

Researchers steal proprietary LLM inference chains via cross-model compatibility of encrypted inference blocks