Translation in progress Some entries still include text in the original language.
Oct 13 · Tuesday
All papers Found 20 papers Another 201 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.
Proposing an option channel attack that renames option labels so that agent guardrails allow prohibited operations
Published Oct 9
Evaluated models LAYA-TD , LAYA-EN , LAYA-ML and others5 Abstract The 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.
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 评估应用决策层的输入注入与隐私推断风险
Published Oct 4
Evaluated models Jev , NanoJev Abstract Jev 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。
提出 ASA 与 AttestMCP，度量并降低 LLM 对输入序列扰动的敏感
Published Oct 2
Evaluated models Gemma-3-27B-Instruct , Gemma-3-4B-Instruct , Llama-3.2-3B-Instruct and others12 提出 imMRAG 攻击，从多模态 RAG 私有图像库提取内容
Published Oct 1
Evaluated models Lumina-mGPT , Gemini 2.5 Flash Image Preview , CLIP ViT-B/16 and others5 提出 ALIBI，向恶意二进制注入不执行掩护叙事使分析器误判为良性
Published Sep 17
Evaluated models Gemini 2.5 Pro , GPT-5.5 Pro , Claude Opus 4.7 Abstract ALIBI 用二进制内的掩护叙事重释可疑静态证据，Gemini 上翻转多，另两模型主要为降级。ALIBI 是针对 LLM 恶意软件分诊的语义掩护叙事攻击：不覆盖系统指令，而把仍可见的可疑静态证据重释成良性安全产品的预期行为。
提出 CIG-MIA，用上下文信息增益推断 RAG 知识库成员身份
Published Sep 14
Evaluated models Llama-3.1-8B-Instruct , Mistral-7B-Instruct-v0.3 , Llama-2-13b-chat-hf and others6 Abstract CIG-MIA 用注入前后的答案似然差推断 RAG 库成员，灰盒可分性高于所列基线。CIG-MIA 是针对 RAG 外部知识库的成员推断攻击，同时覆盖灰盒和纯文本黑盒。
提出 OBELISK 三级流水线，权衡恶意软件检测的准确率、开销与鲁棒性
Published Sep 8
Evaluated models GBDT-ALL , GBDT-YARA , NEBULA-ALL and others4 Abstract 三级过滤以不大的检测损失换速度，同一机制扩大攻击面，部署应按约束选配置。OBELISK 把 Windows PE 恶意软件检测做成 YARA、EMBER-GBDT 与 Nebula 的顺序流水线，用来同时看检出、计算开销和对抗鲁棒性。
Proposes SHAQ, which replaces document embeddings with shadow query embeddings to defend against embedding inversion.
Published Sep 4
Evaluated models GTR-T5 encoder , DPR Abstract SHAQ replaces document embeddings with shadow query embeddings.SHAQ is a defense against EIA for cloud vector databases: instead of directly modifying document embeddings, it replaces them with shadow query embeddings.
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 LeakGauge，用行为后缀的 prefill 对数概率检测上下文泄露
Published Aug 18
Evaluated models Gemma-4-E4B-it , Gemma-4-31B-it , Qwen2.5-14B-Instruct and others11 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。
提出分片监督，将评判标准分配给独立调用以抵御展示攻击
Published Aug 6
Evaluated models Claude Opus 4.8 , Claude Sonnet 4.6 , Claude Opus 4.6 and others4 Abstract 论文揭示决策负载是大模型监督的独立瓶颈，提出分片与辩论机制有效分离注意力与平衡证据，在多项高难度基准上实现稳健评判。本文系统研究了大语言模型在执行多标准监督与评判时的决策负载瓶颈及其安全风险。
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。
提出嵌入推断攻击 EIA，在黑盒 IR 中识别所用嵌入模型
Published Jul 1
Evaluated models e5-small , e5-small-v2 , e5-base and others19 Abstract EIA 用无序文档集的差异从候选中识别嵌入模型，并给出 reranker 与阈值下的计数。EIA 是在黑盒 IR 上识别所用嵌入模型的攻击。 。
Propose soft prompt defense AREA, re-anchoring attention to mitigate system prompt leakage
Published Jun 17
Evaluated models Llama-2-7B-chat-hf , Llama-3.1-8B-Instruct , Mistral-7B-Instruct-v0.3 and others8 Abstract The authors measure commercial prompt leakage, propose attention drift, and use AREA to strike a trade-off between usability and leakage resistance.The authors study system prompt leakage in real LLM applications and propose AREA, a mitigation method that does not modify weights.
提出 MEntA，用非模板查询推断文档是否属于 RAG 检索库
Published May 23
Evaluated models Phi4-14B , Llama3.1-8B , CommandR-7B and others5 提出 SWhisper，用近超声波向语音驱动 LLM 隐蔽注入越狱提示
Published Mar 14
Evaluated models Meta-Llama-3.1-8B-Instruct , Gemma-3-4B , Qwen2.5-7B-Instruct and others9 提出改评分标准的搜索攻击，使 LLM 评判者在基准达标下发生偏好漂移
Published Feb 14
Evaluated models Qwen3-14B , Gemma-3-27B-it , DeepSeek-V3 and others8 Abstract RIPD 指 benchmark 合规的 rubric 修改仍可使 judge 偏好在 target 域定向漂移，并经 DPO 传入策略。作者把自然语言 rubric 视为 LLM judge 的可编辑控制接口，并称只靠 benchmark 验证不足以稳住 target 域偏好。
提出 DECEIVE-AFC，以声明改写诱导搜索增强事实核查误判
Published Jan 31
Evaluated models HiSS , LEMMA , DEFAME and others4 You have reached the end