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

Oct 12 · Monday

All papers

Found 16 papers

Filter4

Select any value within a category; different categories must all match. Counts include the other filters you have selected.Another 233papers have no tags yet and are not included in filtered results.

Attack surface / entry point：Tool responses Attack surface / entry point：Memory writes Attack surface / entry point：Weights and gradients System component：Guardrails and judges Clear all

Another 233papers do not have category tags yet and are not included in filtered results.

AttacksarXiv 2610.12292Oct 9
The paper proposes an option channel attack: changing one option label can make an agent guardrail allow a prohibited operation
Proposing an option channel attack that renames option labels so that agent guardrails allow prohibited operations
arXiv
2610.12292
Published
Oct 9
Layer
Application layer
Setting
LLM applications
Attacker
Black-box
Evaluated models
LAYA-TD, LAYA-EN, LAYA-ML and others5
- AttacksDetection evasion
- ComponentGuardrails and judges
AbstractThe authors' evaluation indicates that typed models should not serve as final safety arbiters; they are susceptible to off-policy text and option name manipulation, and confidence filtering cannot resist attacks.
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
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
Evaluation and benchmarksarXiv 2610.03448Oct 2
Study retests 15 prompt injection detectors: benchmark scores poorly predict agent deployment performance
重测提示注入检测器在 Agent 工具输出上的检出与任务阻断
arXiv
2610.03448
Published
Oct 2
Layer
Application layer
Setting
AI Agent
Evaluated models
Horizon-Labs, Wolf Defender, Prismor-1.5B and others15
- DefensesDetection and filtering
- AttacksPrompt injection
- ComponentGuardrails and judges
Abstract论文揭示公开基准无法反映智能体检测器表现，检测能力源于输入格式相似而非通用防御，建议基于工具输出与低误报评测。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.35886Sep 27
Agentic Commerce Bench：衡量会花钱的智能体的欺诈检测能力
构建支付欺诈基准 ACB，评测花钱智能体的欺诈检测
arXiv
2609.35886
Published
Sep 27
Layer
Application layer
Setting
AI Agent
Evaluated models
gordonguard, pattern judge
- DefensesDetection and filtering
- RiskDangerous agent actions
- ComponentGuardrails and judges
AbstractACB 把支付欺诈与智能体操纵分开测量，作者强调观察权限、误报预算和微支付复核成本。
ACB 是一套智能体支付欺诈基准与检测参考栈，目标是衡量钱是否被不当提取，而不仅是智能体是否接受了恶意指令。
Paper analysisRead analysis
DefensesarXiv 2609.33039Sep 27
TACIT: detecting agent trajectory harm from LLM internal states
提出 TACIT，从冻结内部状态检测智能体有害轨迹
arXiv
2609.33039
Published
Sep 27
Layer
Application layer
Setting
AI Agent
Evaluated models
Qwen3-4B, Llama-3.1-8B, Qwen3-4B-Instruct and others4
- DefensesDetection and filtering
- AttacksPrompt injection
- ComponentGuardrails and judges
Abstract提出基于内部表征的智能体轨迹安全检测方法 TACIT，揭示工具风险表征特性并在多项基准上提升检测表现与效率。
Paper analysisRead analysis
Evaluation and benchmarksarXiv 2609.29429Sep 24
Jev 被用于零样本检测十类对齐失败，中位 AUROC 达 0.886
提出 RLCDAlignBench，零样本检测多种对齐失败
arXiv
2609.29429
Published
Sep 24
Layer
Prompt layer
Setting
Models and APIs
Evaluated models
Jev (jev-1.13.0)
- DefensesDetection and filtering
- AttacksJailbreaks
- RiskMisalignment and value divergence
- ComponentGuardrails and judges
Abstract一次通用问题已能给多数对齐失败排序，大的误差来自状态和标签。
RLCDAlignBench 测量用强化学习做校准决策的 Jev，能否当作对齐失败的零样本检测器。
Paper analysisRead analysis
DefensesarXiv 2609.10707Sep 10
面向 AI-SOC 的神经符号防御框架：用 SIEM 解码器与 NeMo Guardrails 阻断日志投毒提示注入
提出神经符号纵深架构，阻断 AI-SOC 的日志投毒与间接提示注入
arXiv
2609.10707
Published
Sep 10
Layer
Application layer
Setting
LLM applications
Evaluated models
Gemma 3 (12B)
- DefensesDetection and filtering
- AttacksPrompt injection
- ComponentGuardrails and judges
Paper analysisRead analysis
AttacksarXiv 2608.03844Aug 4
MAFIA：针对带审计 LLM Agent 的查询式记忆投毒攻击
提出 MAFIA，以探测放置与事实伪装对带审计 Agent 做查询式记忆投毒
arXiv
2608.03844
Published
Aug 4
Layer
Application layer
Setting
AI Agent
Evaluated models
gpt-5.4-mini, Gemini, ProtectAI and others8
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentMemory
Paper analysisRead analysis
DefensesarXiv 2607.25297Jul 28
MTGuard: Hybrid Analysis for Secure MCP Tool Calls in LLM Agents
Proposes MTGuard, which blocks unsafe tool calls by comparing MCP call declarations with execution traces.
arXiv
2607.25297
Published
Jul 28
Layer
Application layer
Setting
AI Agent
Evaluated models
DeepSeek-V4-Flash, DeepSeek-V4-Pro, GPT-5.6 Luna
- DefensesMonitoring and auditing
- AttacksPoisoning and backdoors
- ComponentMCP and skills
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 2604.03890Apr 5
From prompts to physical actions: structured backdoor attacks against LLM-mediated robot control systems
Propose a structured JSON backdoor that triggers malicious actions in robot control via a poisoned LoRA
arXiv
2604.03890
Published
Apr 5
Layer
Training and data layer
Setting
Embodied systems and robotics, LLM applications
Evaluated models
Llama-3.1-8B-Instruct, Gemma-2-9B-IT, DeepSeek-R1-Distill-Llama-8B and others4
- AttacksPoisoning and backdoors
- TechniqueDormant triggers
- ComponentFine-tuning and open weights
Abstract推理层后门传不到执行。
Paper analysisRead analysis
DefensesarXiv 2603.14707Mar 16
Visual Confusion Agents: Exploiting and Defending Against Perception Failures in Computer-Use Agents
Proposes out-of-process dual-channel comparison classification to intercept erroneous clicks by computer-use agents
arXiv
2603.14707
Published
Mar 16
Layer
Application layer
Setting
Browser and computer use
Evaluated models
multi-modal-embed-small
- DefensesDetection and filtering
- AttacksPrompt injection
- RiskDangerous agent actions
- ComponentGuardrails and judges
AbstractPerception errors can be exploited by swapping screen pixels.
The authors frame CUA's perception failure as a security vulnerability, visual confused deputy: the agent authorizes click(x,y) based on a misread screen, and the actual landing point differs from the element it believes it is clicking.
Paper analysisRead analysis
AttacksarXiv 2602.01658Feb 2
Efficient Adversarial Attacks on High-Dimensional Offline Bandits
提出三类攻击，训练前扰动奖励模型以劫持离线 bandit
arXiv
2602.01658
Published
Feb 2
Layer
Training and data layer
Setting
Models and APIs
Evaluated models
Image Reward, Aesthetic Model
- AttacksPoisoning and backdoors
- ComponentGuardrails and judges
Abstract公开奖励权重可被小扰动劫持离线 bandit，打乱日志只部分缓解。
作者研究离线 bandit 在奖励模型被事先改权重时，还会不会选出最优臂。
Paper analysisRead analysis
DefensesarXiv 2601.10156Jan 15
ToolSafe 提出步骤级护栏 TS-Guard 与 TS-Flow，降低 Agent 有害工具调用
提出 TS-Guard 与 TS-Flow，在执行前检测并引导 Agent 修正有害工具调用
arXiv
2601.10156
Published
Jan 15
Layer
Application layer
Setting
AI Agent
Evaluated models
GPT-4o, TS-Guard, Qwen2.5-14B-Instruct
- DefensesDetection and filtering
- AttacksJailbreaks
- ComponentGuardrails and judges
AbstractTS-Bench 标步骤级风险，TS-Guard 用多任务 RL 做执行前判定，TS-Flow 用反馈代替直接中止。
ToolSafe 研究 LLM 智能体在工具执行前的步骤级安全：基准 TS-Bench、护栏 TS-Guard，以及反馈驱动框架 TS-Flow。
Paper analysisRead analysis
DefensesarXiv 2511.20597Nov 26, 2025
BrowseSafe：研究者构建浏览器 Agent 提示注入基准并评测现有防御
提出 BrowseSafe，在工具输出边界检测并拦截浏览器 Agent 提示注入
arXiv
2511.20597
Published
Nov 26, 2025
Layer
Application layer
Setting
web agent
Evaluated models
Qwen3-30B-A3B-Instruct-2507, Qwen3-8B, PromptGuard-2 (22M) and others10
- DefensesDetection and filtering
- AttacksPrompt injection
- ComponentGuardrails and judges
Paper analysisRead analysis

You have reached the end

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

Attack surface / entry point

Layer

System component

The paper proposes an option channel attack: changing one option label can make an agent guardrail allow a prohibited operation

DEFER1 多智能体防御研究报告受控测试中的攻击成功率下降

SecJev：为 System One 决策模型引入安全专业知识

Study retests 15 prompt injection detectors: benchmark scores poorly predict agent deployment performance

Agentic Commerce Bench：衡量会花钱的智能体的欺诈检测能力

TACIT: detecting agent trajectory harm from LLM internal states

Jev 被用于零样本检测十类对齐失败，中位 AUROC 达 0.886

面向 AI-SOC 的神经符号防御框架：用 SIEM 解码器与 NeMo Guardrails 阻断日志投毒提示注入

MAFIA：针对带审计 LLM Agent 的查询式记忆投毒攻击

MTGuard: Hybrid Analysis for Secure MCP Tool Calls in LLM Agents

研究者提出 Context-Fractured Decomposition 攻击

From prompts to physical actions: structured backdoor attacks against LLM-mediated robot control systems

Visual Confusion Agents: Exploiting and Defending Against Perception Failures in Computer-Use Agents

Efficient Adversarial Attacks on High-Dimensional Offline Bandits

ToolSafe 提出步骤级护栏 TS-Guard 与 TS-Flow，降低 Agent 有害工具调用

BrowseSafe：研究者构建浏览器 Agent 提示注入基准并评测现有防御