Translation in progress Some entries still include text in the original language.
Oct 13 · Tuesday
Agent safety · Papers Found 16 papers “
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。
评测编码智能体 harness 安全机制在对抗任务中的防护效果
Published Oct 6 提出 VAL 框架，并用确认门与参数沙箱约束智能体高风险工具调用
Published Oct 3
Evaluated models DeepSeek-chat , Meta Llama 3.1 8B , Kimi 为 Agent 提出资源层授权系统 Pincer，用数字分身执行动态最小权限
Published Oct 1
Evaluated models Claude Haiku 4.5 , Muse Spark 1.2 Contributor Abstract Pincer 在资源层以三票合取做授权。Pincer 是加在现成 coding agent 上的资源层授权防御：用只受用户侧信息影响的 digital twin 做动态最小权限，并与 tool-call 层防御并存。要处理的是用户维护策略带来的权限蠕变和授权疲劳，以及不学习用户策略的 tool-call 分类器。作者引述工作称 auto mode 会放行多数不安全操作，并认为推理代码或复杂 shell 的效果在这一层不实际。
提出 FlowReview，用授权配对与确定性提交门控控制多智能体组合信息流
Published Sep 30
Evaluated models Sonnet 4.5 , Haiku 4.5 , Qwen3-32B and others5 Abstract 配对标准要求拒绝禁止使用并完成授权使用，FlowReview 把三项审查接到执行。这篇工作把多智能体安全写成一件必须同时成立的事：拦住组合信息所启用的禁止使用，并完成同一场景里任务所需的授权使用。
提出 Tracekit，审计编码 Agent 的意图、推理与动作并检测账本篡改
Published Sep 29
Evaluated models Claude Code (claude -p, version 2.1) , 独立评审（同一 CLI，无工具） Abstract Tracekit 把意图、自述和动作写入可锚定的哈希链。Tracekit 是无第三方依赖的编码智能体审计层：把人类请求、模型自述和已执行动作并排写入可外部锚定的哈希链，执行前用正则门，事后用规则和独立模型互查。
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 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。
用 FARSIGHT 方案级评测金融 LLM 交易 Agent 的稳健性与安全失败
Published Sep 17
Evaluated models ChatGPT Deep Research 用任务绑定令牌和策略预言机拦住多智能体流水线的未授权执行
Published Sep 15
Evaluated models gemma4:31b-cloud Abstract 结构性授权拦住了记忆投毒的未授权执行，持密钥时承重的是策略预言机。作者在一条会改写沙箱邮件库的四智能体流水线上，检查低权限失陷会不会变成未授权的真实动作。
提出 ATF 与 NTR，过滤或恢复工具集以阻止 Agent 调用攻击工具
Published Sep 14
Evaluated models Gemma2-9B , Qwen2-7B , LLaMA3-8B and others7 提出 AgentKernel，用四支柱强制中介 Agent 的输入、工具与记忆
Published Aug 29 Abstract AgentKernel 是四支柱智能体 OS 设计，安全论证未附带实测数字。AgentKernel 是把安全做成组织原则的智能体 OS 架构，不是带跑分的攻防评测。
构建自演化基准 ActBench，评测协作智能体的行为安全
Published Aug 10
Evaluated models Claude-Opus-4.8 , Claude-Sonnet-4.6 , GPT-5.5 and others15 Abstract ActBench 通过自演化搜索与双证据重构评测轨迹行为安全，揭示基础模型决策差异主导了协同智能体的操作风险。
提出 TMA-NM，以写时来源绑定抵御 Agent 长期记忆投毒
Published Jun 23
Evaluated models gpt-5-chat , gpt-4o-mini , claude-opus-4.1 and others8 Abstract TMA-NM 用来源绑定的行动权限挡住记忆洗白。TMA-NM 把 LLM 智能体长期记忆的行动权限绑到写时来源，用来挡住跨会话记忆投毒。
建模并利用多智能体网络中子智能体生成时的记忆继承漏洞
Published May 9
Evaluated models MiniMax M2.5 , Llama-4-Maverick-17B-128E-Instruct-FP8 , Qwen3.5-Plus and others5 Abstract 形式化子智能体继承，在 OpenClaw 上演示四类编排层漏洞并提出防御。作者对多智能体网络里的子智能体继承 做形式化分析，并用开源框架上的 PoC 演示编排层如何让单点妥协继续扩散。
Proposes a defense-in-depth aligned with the five stages to mitigate cross-stage threats to autonomous agents
Published Mar 12 Abstract This is a lifecycle threat analysis and layered mitigation design for OpenClaw, with the cases serving as qualitative evidence.This is a lifecycle threat analysis of OpenClaw, accompanied by a set of layered mitigation designs. The problem addressed is: when autonomous agents connect instant messaging, persistent memory, third-party plugins, and high-privilege execution, attacks are no longer limited to single-turn jailbreaks but can compound across multiple stages from initialization to execution. The authors argue that defenses hardening only individual interfaces cannot stop such cross-temporal attacks.
You have reached the end