Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 22 papers Another 240 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.
提出 BRANCH 方法，用分支搜索绕过多扫描器护栏
Published Oct 8
Evaluated models Meta Llama Guard 3 8B , Meta Llama Guard 7B , Meta Llama Guard 3 1B and others18 提出 PyCache Trap，用字节码缓存替换绕过 Agent 技能扫描器
Published Oct 7
Evaluated models CISCO , HSS , AUDITOR and others8 Abstract 论文剖析智能体技能审查与执行脱节盲区，提出 PyCache Trap 攻击并构建基于可信复现的 EAV 防御。这篇论文研究了智能体技能（Agent Skills）安全扫描中的“审查–执行脱节”问题，并提出了对应的攻击验证与防御方案。
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
Published Oct 7
Evaluated models Qwen2.5-Coder-7B , Qwen3-Coder-30B , Devstral-Small-2-24B and others19 Abstract PackHallu 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。
提出 CORSA，按任务簇优化技能注入的检索与执行
Published Oct 6
Evaluated models GPT-5.4 , DeepSeek-V4-Pro , Qwen3.8-27B and others5 提出 CoordPoison，解耦技能投毒的借口与动作
Published Sep 30
Evaluated models DeepSeek-V4-Flash , GLM-5.3-Flash , Claude-Sonnet-4.6 and others5 Abstract CoordPoison 外置借口。CoordPoison 是一种 Skill 供应链投毒方法，把“为何执行”和“执行什么”拆到两个 Skill。 。
提出 LENS，构造集合级 RAG 语料投毒以导向预设答案
Published Sep 28
Evaluated models Qwen3.6-35B-A3B , Llama-3.1-70B-Instruct , DeepSeek-V4-Pro 提出 SkillDRE，用预执行与运行时双阶段反馈演化恶意 Agent 技能
Published Sep 26
Evaluated models DeepSeek-V4-Pro , GLM-5.2 , Qwen3.5-397B-A17B and others4 提出 VerTox，用可验证奖励强化学习生成误导文档并投毒检索语料
Published Sep 1
Evaluated models SimLM-base , BGE-base , Cohere-embed-english-v3.0 and others8 Abstract VerTox 以可验证奖励微调小 LLM，生成能抬高排序并腐蚀事实的投毒文档。VerTox 用可验证奖励把小 LLM 微调成语料投毒生成器，用来探测神经排序器在注入文档下的暴露面。
提出 RedEvoAgent，演化攻击技能对黑盒智能体做自动红队
Published Aug 28
Evaluated models MiniMax-M2.5 , DeepSeek-V4-Flash , Qwen3.5-35B Abstract RedEvoAgent 用验证棘轮演化可读攻击技能，在两种产品级 harness 上多数设置高于固定工具与红队基线。RedEvoAgent 是面向产品级 black-box 智能体的自动红队方法，用一份可演化、人类可读的攻击技能编排多个越狱工具。 。
提出 SKILLBLOAT，改写技能文档放大编码 Agent token 消耗
Published Aug 22
Evaluated models GLM-4.7-Flash , GLM-5 , GPT-5.4 Mini and others4 提出 ColluSkill，用跨技能组合绕过 Agent 技能扫描器
Published Aug 10
Evaluated models GPT-5.5 , DeepSeek-V4-Pro , GLM-5.2 Abstract 论文揭示了单技能孤立扫描的盲区，提出多技能链攻击 ColluSkill 与上下文防御 ChainGuard 并验证了有效性。
提出 ElasticBack，在单个 Agent 技能中植入条件后门
Published Aug 10
Evaluated models GLM-5.2 , MiniMax-M3 , GPT-5.4 and others4 Abstract ElasticBack 用冻结规则加遗传搜索演化 trigger，在单 skill 上做条件后门。ElasticBack 是放在 LLM agent 单个 skill 里的条件后门攻击。 。
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。
提出对抗重构，仅改表述以拉高 AI 审稿分数
Published Jun 11
Evaluated models Claude Sonnet 4 , Claude Sonnet 4.5 , GPT-5-mini Abstract 揭示 AI 审稿系统易受纯表述重构操纵，建议在部署前解耦科学与写作评价。
提出 DiscourseFlip，用有限文档投毒操纵黑盒 RAG 的话语级观点
Published May 31
Evaluated models Llama3.1-8B-Instruct , Qwen3-8B-Instruct , BGE and others5 Abstract DiscourseFlip 按语义图分配投毒预算。DiscourseFlip 是针对黑盒 RAG 的话语级观点操纵。作者要处理的缺口是：只打单条查询或主题局部查询时，覆盖窄，目标受审查时容易被隔离，也难沿相关话题传播。攻击者不能看检索器或 LLM 内部，也不能改提示词模板，只能在文档数、生成 token 和 SEO 编辑距离限制下注入少量文档。
提出 TRACE 框架，用分解与可演化多轮策略诱导智能体执行有害工作流
Published May 29
Evaluated models Codex , Claude Code , GPT-5.5 and others6 提出 CITA 三阶段框架，生成可绕过毒性检测器的中文隐式毒性文本
Published May 21
Evaluated models Tencent , Baidu , Gemini 3.1 Pro and others11 Proposes AdversarialCoT single-document retrieval poisoning to manipulate RAG reasoning answers
Published Apr 14
Evaluated models DeepSeek-R1 , Qwen3 , GLM4.5 and others5 提出 SkillJect，自动生成投毒 skill 诱导代码 Agent 执行隐藏脚本
Published Feb 16
Evaluated models Claude-4.6-Sonnet , GPT-5-mini , GLM-4.7 and others10 Abstract SkillJect 用双通道和闭环精炼投毒 skill。SkillJect 是针对可安装 agent skill 的自动化提示注入框架。作者要解决的是：公开或手工放入 skill 的恶意指令常常被拒绝或忽略，而 skill 一旦安装就会被反复当作可信指导加载。
提出 Confundo，微调投毒生成器为 RAG 生成抗分块的鲁棒投毒内容
Published Feb 6
Evaluated models Qwen3-8b , Llama3-8b , DeepSeek and others8 Abstract Confundo uses a fine-tuned generator for RAG poisoning; the authors report attack results on three objective types: factual, opinion, and hallucination.Confundo turns RAG poisoning into fine-tuning: a small model generates short poison texts for target questions, which are then placed into documents that will be ingested by the knowledge base. It targets the fact that real systems chunk documents and users do not copy preset questions verbatim, weakening specialized poisoning. The attacker has no internal knowledge of the target system and cannot query it, only prepare documents to be ingested.