Translation in progress Some entries still include text in the original language.
Oct 12 · Monday
All papers Found 21 papers Another 233 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.
重测并扩展对 AI Agent 的提示注入，并用 Agent Card 欺骗操纵 A2A 选择
Published Oct 8
Evaluated models Ollama Mistral 0.9.5 , llama3-8b-8192 , gemini-2.0-flash-lite and others5 Abstract 单智能体过滤与分类可被格式或措辞绕过，A2A 选卡还会偏向名称和描述。作者在自建的 LangChain、AutoGen 和 A2A 系统上重测提示注入，并改 Agent Card 看选择是否偏移。摘要称试验主要凸显单智能体弱点，同时把攻击扩到多智能体。
提出利用路由元数据推断敏感话题的侧信道攻击
Published Oct 7
Evaluated models RouteLLM (BERT router) , notdiamond-0001 , Vela (vLLM Semantic Router) Abstract 论文揭示并测量了 LLM 路由决策在无文本日志中的敏感话题泄露，验证了按类别公平性防御的有效性与边界。
提出 CredLeak-Bench，评测 Agent 面对钓鱼的凭据泄露与恢复
Published Oct 6
Evaluated models Claude Sonnet 5 , Claude Opus 4.8 , Gemini 3.6 Flash and others7 提出 TRANSCOPE，用 CPU 硬件计数器推断训练集成员
Published Oct 6
Evaluated models BERT (4.4M-340M) , ViT (22M-307M) , Pythia (70M-2.8B) and others5 Abstract TRANSCOPE 揭示了分词器在嵌入表引发的硬件访问局部性差异，证实微架构计数器可突破恒定时间与置信度遮蔽防御。
提出 DP-MixMin，私有学习公开数据混合并做差分隐私微调
Published Oct 6
Evaluated models ResNet-18 , GPT-2 Medium Abstract DP-MixMin 私下学公开混合再 DP 微调，胸片 macro AUC 与邮件困惑度优于均匀混合。DP-MixMin 是一条端到端的差分隐私训练流程：先私下决定多个公开数据集的预训练混合，再预训练并在敏感数据上 DP 微调。作者称这是首个这样的流程。
提出 UnAct，仅以前向激活干预从分类器中遗忘指定类别
Published Oct 3
Evaluated models ResNet-18 , ViT-B/16 Abstract UnAct 用前向激活做类别遗忘。UnAct 是一种事后类别遗忘：删除请求可以只有少量该类图像，且没有标签和保留数据。它不用梯度，只对遗忘图像做前向。编辑对象是分类器所读的后期单元。
提出 ReGap，用自生成伪监督微调恢复模型中的隐私关联
Published Oct 1
Evaluated models GPT-2 Small , GPT-2 Medium , GPT-2 Large and others6 Abstract 提出无真实私有监督的恢复攻击 ReGap，实验显示微调可通过自生成数据唤醒模型潜在隐私关联，突显后适应隐私风险。
提出 Intervention Score 与 DIR-R，按干预效果选择并修订遗忘参数支持
Published Sep 29
Evaluated models Llama 3 8B , Gemma 2 2B , Qwen2.5-7B and others4 Abstract 作者把局部化遗忘拆成初始干预选择和检查点修订，并在三个设置里分别给证据和边界。Storage Is Not Strategy 研究局部化 LLM 遗忘里两个常被混在一起的决定：一开始该更新哪些参数，以及优化改变模型之后这个选择要不要改。
Published Sep 28
Evaluated models E-Branchformer Abstract 梯度上升类方法更接近简单评测下的重训参照，复杂方法会被知情攻击认出，多说话人更差。
提出主题锚点投毒，放大众包微调后的专有指令提取
Published Sep 28
Evaluated models Qwen2.5-7B-Instruct , Qwen2.5-14B-Instruct , Llama-3.1-8B-Instruct and others4 Abstract 论文提出基于主题锚点的数据中毒攻击，证明低比例的中毒样本可在黑盒条件下隐蔽放大微调指令泄漏。
提出 PMA，用流形对齐从 E2EO 混淆向量恢复明文
Published Sep 7
Evaluated models BERT-base , RoBERTa-base , T5-base and others6 Propose GIFT, isolating user data in shared LLM serving using encryption and GPU information flow tracking
Published Aug 26
Evaluated models Qwen2.5-14B , Qwen2.5-32B , Qwen2.5-72B and others6 Abstract GIFT uses cryptographic isolation plus GPU information flow rules to enforce user data isolation on vLLM and DistServe, and reports throughput overhead.GIFT is the authors' user data isolation system for shared LLM serving, with GIFT-CC enabling confidential computing. In the conclusion, the authors state it is the first IFT design for this scenario.
提出 AgentFlow 流策略语言与运行时框架，约束智能体跨步数据流
Published Aug 24 Proposing KeyPooling to measure where API relays break prompt cache isolation
Published Aug 18
Evaluated models grok-4-fast-non-reasoning , grok-4-1-fast-non-reasoning , grok-3 and others11 Abstract Shared upstream credentials place relay customers in the same cache domain; preserving identity or separate principals can separate them.KeyPooling measures where LLM API relays drop customer identity, causing prompt caches to no longer be isolated per downstream customer.
提出 LoginTrap，以网页弹窗诱导 Web Agent 提交敏感信息
Published Aug 5
Evaluated models GPT-4o , Gemini 3 Flash , Claude Sonnet 4 and others7 Abstract 揭示网络智能体在黑盒网页下的登录诱导风险，利用模糊测试生成话术并跨骨干和架构验证了隐私泄露威胁。
提出 TC-UAP，用身份级多帧扰动保护视频免被定制复现身份
Published Jul 15
Evaluated models LTX-2.3 , Wan2.2-5B Abstract TC-UAP 以可复用的时间一致扰动干扰两类定制共用的 VAE 隐空间。作者要解决的是：发布前如何给视频加不可见扰动，使片段既不能被拿去微调视频扩散模型，也不能被抽帧做 image-to-video。
提出面向智能体浏览器的 Web 威胁模型并复现混淆攻击
Published May 7
Evaluated models GPT-5 , Claude-Sonnet-4.6 , Kimi-k2p5 and others8 Abstract 作者提出面向智能体浏览器的 Web 威胁模型与攻击分类，测试显示传统混淆攻击可复现且能绕过现有防御。这篇论文研究了智能体浏览器在面临传统 Web 攻击者时的安全性问题，指出当前针对提示词注入的安全假设无法抵御基于正常 Web 交互模式的混淆攻击。
提出 ClawSafety，评测个人智能体读到对抗工件后是否执行有害动作
Published Apr 2
Evaluated models Claude Sonnet 4.6 , GPT-5.1 , Gemini 2.5 Pro and others5 Proposes PrefixWall, which selectively isolates shared prefixes to mitigate timing side channels
Published Mar 11
Evaluated models gemma-3-4b-it , llava-onevision-qwen2-0.5b-ov-hf , llava-onevision-qwen2-7b-ov-chat-hf and others9 Abstract PrefixWall uses cache metadata to selectively cut off cross-user prefix continuation, preserving APC while limiting the TTFT side channel.PrefixWall is a system-level defense for multi-tenant LLM services that limits the TTFT side channel of Automatic Prefix Caching while preserving prefix reuse.
Proposing OPTILEAK, reconstructing others' prompts token-by-token via a KV cache side channel
Published Feb 24
Evaluated models Qwen-2.5-3B-Instruct , Qwen-2.5-7B-Instruct , Qwen-2.5-14B-Instruct and others4 Abstract OPTILEAK uses SFT and hard-token DPO to reduce the number of requests per token; the authors claim the risk of cache prompt leakage is higher than previously reported.