Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 21 papers Another 232 papers do not have category tags yet and are not included in filtered results.
“
About this collection Collection in progress We are building a large AI safety paper collection to help researchers find work related to their interests. Papers are collected automatically andfiltered only for relevance to AI safety , without judging quality, novelty or reading value. The default “Annotated first” shows papers with tags or published paper analyses first. Each group is ordered by original publication date, newest first, with undated papers last. “Newest first” uses publication dates only. This order is not a quality ranking.
One-line summaries, tags and paper analyses are prepared by AI from the papers. Language models may have systematic biases and produce hallucinations.We plan to carry out human annotation and review to reduce errors in this content.
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
Published Oct 6
Evaluated models Qwen3-235B-A22B-Instruct-2507 , gpt-oss-120b , Llama-4-Scout-17B-16E-Instruct and others5 Abstract DEFER 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。
提出检测、结构化与 GSPO 流程，降低共识生成的提示注入错位
Published Oct 3
Evaluated models LLaMA-2 7B , LLaMA-3 8B , GPT 3.5 Turbo and others14 Abstract 三章依次做个体偏好预测、政党知识图谱审议，以及共识生成对提示注入的脆弱性与一条防御流程。这是一篇把 LLM 数字孪生放进增强民主的博士论文，分个体表示、政党聚合和共识生成的提示注入三章。
提出 EviLLM，经账户指令或插件向代码生成流程注入漏洞
Published Oct 3
Evaluated models GPT-4o , Llama 3.2 (3B) 提出 MADBench，评测多智能体辩论是吸收还是放大对抗影响
Published Sep 30
Evaluated models gpt-4o , GPT-5 , GPT-6 Astra Abstract MADBench 显示辩论能吸收部分准确率攻击，同时放大未授权读写。MADBench 是一个评测多智能体辩论安全性的基准，用来看攻击是被辩论吸收还是被放大。
提出 VirusCascade，联合操纵商品描述与交互轨迹以劫持推荐智能体协同反思
Published Sep 29
Evaluated models AgentCF , AgentSEQ , AgentRAG and others4 Abstract 作者把反思写回视为推广攻击入口，并用联合操纵和诊断探针研究其持续与传播。VirusCascade 是针对推荐智能体协作反思的 black-box targeted promotion 攻击研究。 攻击者只改商品描述和少量账号的普通交互，试图让目标商品获得更多良性用户曝光。
提出残余权限重放攻击，跨任务复用长期 Agent 的历史授权绕过确认
Published Sep 28
Evaluated models GPT-4.1 , Gemini-3.1-Flash-Lite , Qwen3-14B and others6 提出 ORBIT 框架，评测多智能体防御对合谋等威胁的效果
Published Sep 27
Evaluated models GPT-5.4 , Claude Sonnet 4.6 , Qwen3-235B and others5 Abstract ORBIT 用六层配置比较多智能体防御，逐动作监控不迁移到合谋。
提出 CAVE-BENCH，评测智能体在虚假指责下破坏已完成的正确工作
Published Sep 26
Evaluated models Claude-Sonnet-5 , Claude-Opus-5 , GPT-5.6-Sol and others14 Abstract 论文提出了评估智能体在虚假指责下破坏已有正确成果的 CAVE-BENCH，发现强模型易推翻已知证据，提出了门控缓解方案。
测试编程智能体被诱导或自主篡改自身执行轨迹的行为
Published Sep 25
Evaluated models GPT-5.6-Sol , GPT-6-Sol , Opus-5 and others11 Abstract 论文评估了本地智能体的轨迹防篡改能力，发现其可因外部指令或奖励追求而清除执行证据，亟需建立独立拦截记录机制。这是一项针对本地大模型智能体执行轨迹完整性与安全防篡改边界的实证研究。
提出隐蔽后门攻击 AKRASIA，在推理代码模型提示中植入触发器并伪装思维链
Published Sep 1
Evaluated models Claude-Sonnet-5 , GPT-5.5 , Gemini-3.5-flash and others7 Abstract 论文针对推理代码大模型提出了隐蔽推断期后门攻击 AKRASIA，揭示了模型思维链可被利用进行欺骗性伪装的安全隐患。
提出 ClawSentry 多层监控网关，在技能准入和运行时拦截恶意技能与隐藏指令
Published Aug 21
Evaluated models Codex/GPT-5.4 , Codex/GPT-5.5 , Claude Code/GLM-5.1 and others6 Abstract ClawSentry 用包准入、三级运行时门控和反绕过，压低 skill 攻击并保住干净任务。ClawSentry 是挂在智能体运行时之外的安全监督网关，用来同时盖住 skill 执行的多个生命周期边界，并把已被拒绝的效果记到会话里。
提出 S³，用 guard agent 编排阶段安全技能防御智能体多阶段风险
Published Aug 3
Evaluated models DeepSeek-V4-Pro , DeepSeek-V4-Flash Proposes a static-first validator, CARE, that decides before dispatch whether an agent's shell command may be executed
Published Jul 22 提出 AHA 自动科研循环，发现生产 Agent 的可复用漏洞概念
Published Jul 13
Evaluated models Minimax-M2.7 , Kimi-K2.6 , Deepseek-V4-Pro and others5 Abstract 论文提出基于可证伪科研循环的 AHA 框架，从生产级智能体中挖掘并沉淀漏洞概念图，提升攻击复用率并指导防御加固。AHA 是一项针对生产级 LLM 智能体自动化红队测试的研究，旨在解决传统方法仅保留具体攻击载荷而缺乏成因解释、导致跨版本复用困难且无法指导修复的问题。
提出 Agent MechSuits，对多轮 CLI Agent 做步级检测与子空间转向
Published Jun 22
Evaluated models LLaMA-3.1-8B , Qwen-2.5-7B , Gemma-2-9B Abstract 白盒步级探针加稀疏转向可压低三模型恶意动作，同时提高生成崩溃率。Agent MechSuits 是面向多轮 CLI 智能体的白盒、步级表示防御，并配有可执行轨迹基准 MAS。
提出 MAStrike，用 Shapley 选联盟对多智能体做共谋红队
Published Jun 11
Evaluated models Claude Opus 4.7 , GPT-5.5 , Gemini 3.1 Pro 提出 SciTrace，为科学发现智能体加入跨阶段安全推理与工具轨迹核验
Published Jun 6
Evaluated models Llama-3.1-70B-Instruct , Qwen2.5-72B-Instruct , DeepSeek-V3 and others4 Abstract SciTrace 把安全推理织入科学智能体各阶段，在 SciSafetyBench 上提高安全指标并保持输出质量。SciTrace 是加在科学发现智能体流水线上的安全框架，不是再在各阶段出口放一个独立过滤器。
Proposing a lineage-graph governance layer to suppress error cascades in multi-agent collaboration
Published Mar 4 Abstract A single seed can solidify into a false consensus along a fixed collaboration graph.The authors turn "small errors becoming collective false consensus" in LLM-MAS into a measurable propagation on a graph, and use a plug-in that only changes messages, not topology, to reduce the final infection.
Propose Thought Virus, which propagates preferences along multi-agent messages via subliminal prompts.
Published Feb 23
Evaluated models Qwen2.5-7B-Instruct , Llama-3.1-8B-Instruct Abstract Thought Virus spreads a single-point bias downstream.Thought Virus puts subliminal prompting into multi-agent communication, to see whether a single-point deviation spreads across the network without naming the target concept.
提出 PINA，迭代优化针对导航智能体的提示注入
Published Jan 20
Evaluated models GPT-3.5-turbo , GPT-4 , Balcı et al. [2] Abstract PINA 在黑盒下优化导航提示注入。PINA 是面向基于 LLM 的导航智能体、在黑盒条件下优化提示注入的框架。