深度文章 50

AI/开发情报周报 | 2026年11月20日

近 7 天 339 条情报 + KEV 在野利用 7 条的自动汇总

技术雷达 | 2026年11月:近三个月值得关注的 500 条技术信号

AI、开发者工具、云基础设施与安全四条线的近期信号汇总。

AI/开发情报周报 | 2026年11月14日

近 7 天 340 条情报自动汇总

AI/开发情报周报 | 2026年11月07日

近 7 天 239 条情报自动汇总

AI/开发情报周报 | 2026年10月31日

近 7 天 145 条情报自动汇总

AI/开发情报周报 | 2026年10月25日

近 7 天 69 条情报自动汇总

安全与情报日报 | 2026年10月下旬:KEV 突破 1750 条、AI 基建支出遭减记预警、Anthropic 推迟 IPO

10月下旬三条主线:微软十月CVE清单按利用性排序、KEV在野条目突破1750;Gartner测算AI基建6-7万亿支出面临3.5万亿减记风险;Anthropic IPO推迟至11月。

安全日报 | 2026年9月下旬:边界设备成重灾区,AI 基础设施首次挤进 KEV 前列

9月下旬安全态势复盘:F5 BIG-IP、Check Point、Cisco ISE、WordPress、MikroTik 集中爆零日且确认在野利用;7 枚新增 KEV 中 3 枚指向 AI 基础设施;附优先级修复清单。

GitHub 周报:DeepSeek Harness、OpenViking 与 Agent 生态爆发

本周 GitHub 热点被 AI Agent 基础设施占领:DeepSeek Harness 14 万星破纪录、OpenViking 统一上下文层、Prime Agent、career-ops、ai-memory 等集体上榜。一文速览值得关注的项目。

开源权重模型 vs 闭源 API:企业 AI 选型的分水岭

硅谷数据显示企业推理 token 支出正从闭源流向开源权重模型,DoorDash、Airbnb、Coinbase 等已转用中国开源模型。安全与合规考量成为选型关键变量,本文给出权衡框架。

CVE-2026-73678:MindsDB 未授权 RCE 敲响 AI 基础设施安全警钟

MindsDB 的 Minds Platform 存在 CVSS 10.0 的未授权远程代码执行漏洞,攻击者可无需认证在服务器执行任意命令。AI 数据基础设施成为高危目标,本文给出排查与修复建议。

Agent 成本控制进入会话级:Anthropic 与 AWS 的硬预算实战

8 月 6-7 日,AWS Bedrock AgentCore 与 Anthropic Managed Agents 相继上线会话级硬预算,成本控制从"月度账户上限"下沉到"单次运行"。本文对比各家方案的粒度、边界与超支行为。

供应链攻击的下一个目标:Cursor 与 Claude Desktop 生态

Intel 471 报告指出,攻击者正把投毒目标从传统 npm/PyPI 转向 AI 编程助手生态:恶意 skill、插件与"流行包代理"手法瞄准 Cursor、Claude Desktop 等 Agent 环境。本文解析攻击面与防护建议。

AI 模型定价战:OpenAI 降价 80%、Anthropic 冻结价格,Agent 成本正在重构

7 月底 OpenAI 将 GPT-5.6 Luna 降价 80%、Terra 降价 20%;8 月 Anthropic 宣布 Sonnet 5 的 $2/$10 首发价永久化。开源权重模型的竞争让前沿模型定价进入下行通道,Agent 应用的token经济学被彻底改写。

Prime Agent 与递归语言模型:把 Agent 上下文变成代码

Prime Intellect 开源的 Prime Agent 单日 2319 stars 登顶 GitHub Trending。它让模型持有持久 Python 会话,把文件、命令与子 Agent 全当作代码管理,基于 arXiv:2512.24601 的递归语言模型思想。

OpenViking:字节跳动开源的统一上下文层,Agent 记忆的新范式

字节跳动火山引擎开源 OpenViking,把 Agent 记忆、知识 RAG 与技能合并为单一自进化上下文层,29K stars 单日新增 213。本文拆解其设计思路、适用场景与社区插件生态。

DeepSeek Harness:14 万星的开源 Agent 运行时到底做了什么

DeepSeek 开源 Agent 运行时 DeepSeek Harness 一小时内突破 2 万 stars,四天超 13 万,创 GitHub 史上最快涨星纪录。它用"一切皆插件"的设计统一多种模型平台,本文拆解其架构、执行模式与生态价值。

2026 年 7 月勒索软件态势:新家族崛起,AI 智能体攻击链登场

360 发布 2026 年 7 月勒索软件流行态势分析:ExfilSquad、CRPxO、GSG、Section9、Gammax 五大新家族登场,BrzCrypt 与 Wmansvcs 判定同一团伙。同期 Sysdig 披露首个 AI 智能体勒索攻击链,CVERC 预警 Sorry 勒索软件攻击 Linux 服务器。

AI Agent 框架选型实战 2026:LangGraph、CrewAI、OpenAI Agents SDK 怎么选

2026 年 AI Agent 框架已收敛为 LangGraph、CrewAI、OpenAI Agents SDK 等阵营,它们在架构层面互相反对。本文用同一调研 Agent 分别实现三种方案,剖析状态图、角色团队、Agent 移交三种心智模型的差异,并给出按故障模式选型的决策矩阵。

8月第四周 AI 与开源动态:OpenViking 上下文层与生态展望

字节跳动开源 OpenViking 统一上下文层引发关注。本文展望 Agent 上下文层赛道与 8 月开源生态走向。

GitHub 周报:career-ops 与 AI 求职自动化

AI 求职工具 career-ops 以 6.6 万星上榜。本文盘点 AI 求职与垂直 Agent 工具的开源趋势。

供应链攻击转向 AI 编程助手:企业防护清单

攻击者正把投毒目标转向 AI 编程助手生态:恶意 skill、插件与流行包代理瞄准 Cursor、Claude Desktop。本文给出企业防护清单。

企业 AI 模型选型:开源权重 vs 闭源 API 实操框架

企业推理 token 支出正向开源权重模型迁移,但安全与合规顾虑并存。本文给出一套企业模型选型的实操权衡框架。

DeepSeek Harness 插件开发:Agent 生态的下一站

DeepSeek Harness 以插件化架构破纪录涨星。本文解析 Harness 的插件模型与开发者如何在其中构建生态。

OpenAI 商业化加速:$125 Premium Seats 与 $8 重置配额

OpenAI 为 ChatGPT Business 推出 $125 Premium Seats,并测试 $8 付费重置配额。本文解读 OpenAI 用量变现策略与企业的预算启示。

本地大模型安全部署:Ollama/vLLM 的暴露面收敛

本地推理平台 Ollama、vLLM 的未授权访问是高频风险。本文给出本地大模型部署的安全基线配置清单。

AWS AgentCore 与 Anthropic 会话预算:Agent 成本控制双雄

AWS Bedrock AgentCore 与 Anthropic Managed Agents 先后上线会话级成本控制。本文对比两者的机制、差异与落地建议。

JadePuffer:首个 AI 智能体勒索攻击链拆解

Sysdig 披露 JadePuffer 活动:攻击者利用 AI 智能体自动化侦察、横向移动与勒索部署,被认为是首个以 AI 智能体为核心编排的勒索攻击链。

2026年8月安全前瞻:勒索、供应链与 AI 攻击三条主线

展望 8 月安全态势:勒索软件新家族、供应链投毒扩散至 AI 生态、AI 基础设施攻击。本文给出三条主线的防御优先级。

reverse-skill 登顶 Trending:攻防 Agent 技能路由器

reverse-skill 登上 GitHub Trending 榜首:面向逆向与渗透测试的 Agent 技能路由器,支持 Claude Code、Codex、Cursor。本文解析其设计与安全价值。

GPT-5.6 Luna 降价 80%:AI 推理成本下行的连锁反应

OpenAI 将 GPT-5.6 Luna 价格下调 80%、Terra 下调 20%。本文解读降价动因、对开发者与 Agent 应用的影响。

模型 API 涨价信号:Anthropic/OpenAI 商业化分水岭

OpenAI 与 Anthropic 的定价策略从"补贴扩张"转向"用量变现"。本文解读商业化信号对企业预算与选型的影响。

容器供应链安全:镜像投毒与 SBOM 实践

容器镜像与供应链投毒事件频发。本文解析镜像投毒攻击模式与 SBOM、签名校验等防御实践。

GitHub 周报:DeepSeek Harness 之前,Agent 运行时演化史

在 DeepSeek Harness 爆火之前,Agent 运行时经历了从框架到基础设施的演化。本文梳理 Agent 运行时的发展脉络与未来方向。

OpenAI 硬性支出上限全量上线:API 成本治理新常态

7月23日起 OpenAI API 平台为所有账户提供硬性月度支出上限,超限返回 429。本文解读功能边界与成本治理建议。

勒索软件防御纵深:从防加密到防 Agent

勒索攻击从"加密勒索"演进到"数据窃取+加密+AI编排"。本文给出从防加密到防 Agent 的三层防御框架。

AI 数据层攻击面:MindsDB、向量库与 RAG 安全

MindsDB 未授权 RCE 等事件表明 AI 数据层正成为攻击目标。本文梳理 AI 数据层的攻击面与加固清单。

GitHub 周报:Agent 技能、记忆与上下文层的开源热潮

本周 GitHub Trending 被 Agent 基础设施占领:技能合集、记忆库与上下文层项目扎堆上榜。本文盘点值得关注的仓库与社区信号。

AI安全红队商业化路径:从审计服务到SaaS产品的完整变现指南

系统讲解AI安全红队的商业化路径:从AI安全审计服务($5K-$50K/次)、企业培训($5K-$10K/天)、安全咨询($10K-$200K)到SaaS产品化(月入$18K),涵盖LinkedIn获客、安全会议、漏洞赏金、内容营销四大渠道,含完整Python报告生成器代码和行动清单。

2026年本地部署大语言模型完全指南:用Python运行自己的AI模型

# 2026年本地部署大语言模型完全指南:用Python运行自己的AI模型 ## 为什么要本地部署大语言模型? 每次调用GPT-4、Claude这些商业API,你都在为别人的服务器买单。更关键的是,你的数据——商业计划、代码、客户信息—

社群运营变现教程:从0到万人群主

系统讲解社群运营变现的完整方法论:从社群定位、内容规划、引流策略到付费社群、训练营、企业服务、资源对接四大变现模式。包含Python自动化运营代码和6个月增长路径规划,适合有技术背景的创业者从零搭建高粘性付费社群。

知识付费变现实战教程:把知识变成钱的完整指南,从课程设计到月入5万

知识付费变现完整指南:从付费专栏、训练营、咨询顾问到内容变现四大渠道,涵盖选题策略、AI辅助内容生产、定价阶梯设计、促销玩法和收入预估模型。包含Python自动化选题工具、课程包装清单和推广渠道矩阵,适合有技术背景的开发者和内容创作者从零搭建知识付费业务,实现边际成本趋零的被动收入。

数据标注变现实战教程:AI预标注+Label Studio+团队运营从零到月入5万

数据标注变现实战教程,系统讲解图像/文本/音频/视频四大标注类型、Label Studio开源平台搭建、AI预标注自动化(GPT-4o+YOLO)、三种变现模式(接单外包/自建团队/标注SaaS)和质量控制体系。含完整Python代码示例和月收入预估模型,帮助你从兼职标注员起步,逐步升级为标注团队负责人或SaaS产品开发者。

跨境电商自动化教程2026:AI+Browser Use+数据驱动从零到月入6万,含完整Python代码

本教程讲解如何用AI+Browser Use实现跨境电商全流程自动化,包括Amazon自动选品、多语言AI翻译上架、竞品价格监控和广告智能优化,含完整Python代码和月收入预估模型,预计月入6-8万元。

2026年短视频矩阵运营实战教程:AI批量生产+矩阵发布从零到月入5万,含完整Python代码

2026年短视频矩阵运营完整教程,涵盖抖音/B站/小红书三大平台分析、AI批量生成视频脚本与配音、Browser Use自动化发布、10账号矩阵规划、淘客带货+中视频分成+知识付费三条变现路径。含完整Python代码和月收入预估模型。

AI SEO内容工厂实战教程:用AI批量生产SEO文章实现被动收入完全指南

本教程系统讲解AI SEO内容工厂的完整变现流程:从关键词研究、AI批量生成SEO文章、技术SEO优化到多渠道变现(AdSense、联盟营销、内容付费)。含完整Python代码示例、月收入预估和6-12个月增长路径规划。

AI安全红队概论:OWASP LLM Top 10、攻击面全景与红队工具链实战

本课程系统讲解AI安全红队核心知识:OWASP LLM Top 10逐条解析(Prompt注入、敏感信息泄露、供应链漏洞等),四层攻击面模型(模型层/数据层/应用层/基础设施层),主流红队工具(Garak/PyRIT/ART/TextAttack)对比与实战。含完整Python代码示例、商业AI安全审计服务定价策略,以及使用Garak对本地LLM做首次漏洞扫描的完整实操指南。

AI Agent开发变现实战教程2026:从LangGraph到CrewAI搭建智能客服与研究Agent,含完整Python代码

本教程深入讲解AI Agent开发与商业化变现全流程。涵盖LangGraph智能客服Agent、CrewAI多Agent协作研究系统、销售跟进Agent三大实战案例,对比5大主流Agent框架,提供Agent开发服务、SaaS产品、行业解决方案三种变现模式与月收入预估,含完整可运行Python代码。

淘宝客变现实战教程2026:三大联盟平台注册+佣金结构+推广位创建完整指南,含Python代码

# 第1课:淘客基础与平台注册 > 学完这节课,你将理解淘宝客的商业本质,完成三大联盟平台注册,并创建你的第一个推广位。 --- ## 一、什么是淘宝客(CPS模式详解) ### 1.1 一句话理解淘客 淘宝客 = 帮商家推广商品

2026年Python自动化变现实战教程:从Excel处理到RPA服务月入3万,含完整Python代码

# Python自动化变现教程:从脚本到产品 ## 为什么要学Python自动化 **商业价值:** - 帮企业做自动化:¥3000-20000/项目 - RPA(机器人流程自动化):替代人工操作 - 数据处理服务:清洗/转换/分析 -

实时情报 100

🟢 AI OpenAI makes it much harder for attackers to extract how its latest AI models think OpenAI加强防护措施,防止攻击者提取AI模型内部推理,增强开发者对模型安全性的信任 🟢 AI Bytedance argues most AI training phases are unnecessary and backs it up with a competitive open model Bytedance提出大多数AI训练阶段不必要,并发布竞争性开放模型,对开发者提供新思路 🟢 AI US technology stocks suffer their worst weekly loss since the pandemic as AI doubts pile up 美国科技股遭遇疫情以来最惨跌幅,AI行业面临信任危机,影响开发者投资和信心 🟢 AI New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop Inkling 2 Mini开源模型通过循环思考解决问题,运行效率更高,对开发者提供新的模型选择 🟢 AI Hours after Anthropic claims a Claude proof of the Riemann hypothesis, Stanford releases an all-human competing draft Anthropic声称Claude证明黎曼假设,斯坦福发布全人类证明方案,对数学领域开发者有重要启发 🟢 AI OpenAI shares new prompting tips for GPT-6.2 Astra and Sol to fix brief answers and unnecessary questions OpenAI发布新提示模板,改善GPT-6.2 Astra和Sol的回答质量,提升开发者交互体验 🟢 AI Silicon Valley's growing obsession with AI doomsday is turning parts of the tech industry toward cult-like devotion 佛蒙特州一佛教寺庙欲成硅谷精神权威并创建AI神,对开发者提供新的精神指导方向 🟢 AI Nvidia researchers argue huge AI context windows are a dead end 英伟达研究人员认为超大AI上下文窗口是死胡同,提出更高效的语言模型训练方法。 🟢 AI Hugging Face and OpenRouter join forces to make open-weight models easy to run Hugging Face与OpenRouter合作,使开源模型易于运行,提供更便捷的模型部署方式。 🟢 AI Rumble-2-Flash: an open 0.5B speech model for 40+ languages Hugging Face发布Rumble-2-Flash,开源支持40多种语言的语音模型。 🟢 AI 180B Parameters, 32GB of RAM: Pushing MoE Inference to the Limit with Transformers Hugging Face使用Transformers推动MoE推理极限,支持大规模模型推理。 🟢 AI Time-Series Models in the Streamhouse: A New Frontier and Use Cases Hugging Face探讨时间序列模型在Streamhouse中的应用,提供新颖的用例和解决方案。 🟢 AI Why your fine-tuning job takes longer than the FLOPs say Hugging Face分析为何微调任务耗时长于FLOPs计算,提供优化建议。 🟢 AI DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency Hugging Face发布DaisyRose-27B,推动训练后效率的边界,提供更高效的模型训练方法。 🟢 AI Introducing harness adapters: adapt your favorite model to your harness to improve performance and cost Hugging Face推出Harness适配器,允许开发者将模型适配到特定框架,提升性能和兼容性。 🟢 AI Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup Hugging Face推出Inkling 2 Mini,为本地设置提供高性能模型,提升开发者体验。 🟢 AI New Ask Maps features to help with holiday travel and prep 谷歌地图推出新功能,帮助用户规划假日旅行和准备,提供路线建议、本地预约和保存列表搜索等功能。 🟢 AI Put Gemini to work for you Gemini 任务功能免费开放,助力工作和学习效率提升 🟢 AI The Gemini app is now available on lower-memory Android devices. Gemini 应用适配低内存 Android 设备,提升用户体验 🟢 AI Walter Isaacson on curiosity and creativity in the AI age. Walter Isaacson 讨论 AI 时代的好奇心和创造力 🟢 AI Introducing Tool Retrieval in the Gemini API Gemini API 推出工具检索功能,提升 AI 代理效率 🟢 AI Deploying Daybreak Red Daybreak Red 部署指南,提供监督安全调查的参考架构和安全控制 🟢 AI Get started with Daybreak Blue Daybreak Blue 入门教程,介绍其设置体验和快速上手方法 🟢 AI Muru Muru 讨论 mod_wsgi 的历史和 Python 网页应用的生产环境部署最佳实践 🟢 AI llm-gemini 0.38.1 llm-gemini 0.38.1 发布,优化 Gemini 4 模型推理能力 🟢 AI Token fraud is here and it’s BAD (YouTube) AI 平台遭遇 token 诈骗,开发者需提高警惕 🟢 AI Quoting Levent Alpöge 发现黎曼假设的证明方法,对数学和计算机科学有重要价值 🟢 AI Quoting Ravi Vakil 数学家讨论广义黎曼假设的证明进展和相关研究 🟢 AI See PR scans in action PR 扫描功能演示,展示其在拉取请求工作流中的于代码审查的实践 🟢 AI Explore the Codex Security Plugin Codex 安全插件演示,展示其在代码安全审查工作流中的应用和结果 🟢 AI Can AI help scientists identify promising drug candidates? MedChemBench 评估 AI 模型在药物发现中的性能,帮助科学家识别有前景的候选药物 🟢 AI Hebbia sees improvements in financial research with GPT-6 Astra Hebbia 使用 GPT-6 Astra 提升金融研究效率,在源覆盖率、引用相关性和呈现质量方面取得进展 🟢 AI Combatting adversarial distillation OpenAI 提出对抗性蒸馏防御方法,结合训练、防护、检测和执法保护模型能力,防止被滥用 🟢 AI External sandboxes for coding agents External sandboxes for coding agents I've been exploring new ways to sandbox code that is executed by agents, with a focus on making it safer to try out local models without worrying that a mistake or prompt injection attack might cause the 🟢 AI The PSF end-of-year fundraiser The PSF end-of-year fundraiser With my Python Software Foundation board director hat on: we've launched our end-of-year fundraiser for the non-profit that supports Python development and the Python community. If your company or your work be 🟢 AI Quoting Daniel Stenberg During the time these agents converted 185,000 lines of C code into 340,000 lines of Rust code (I counted), we merged over 300 commits in the actual curl project. We actually improved curl during these a little more than three weeks. A litt 🟢 AI datasette-arrows 0.1a0 datasette-arrows 0.1a0 发布,增强 Datasette 插件功能 🟢 AI llm-gemini 0.38 llm-gemini 0.38 发布,支持 Gemini 4 Neon 模型 🟢 AI PostgreSQL 19 Released PostgreSQL 19 发布,带来数据库性能和功能的更新 🟢 AI Managing frontier AI risk: From training to deployment OpenAI explains how it manages frontier AI risk from training through deployment with model alignment, sandboxing, blocking safeguards, and monitoring. 🟢 AI Inside the Codex sandbox See how Codex runs commands on your machine while enforcing the filesystem and network limits you control across macOS, Linux, and Windows. 🟢 AI Research acceleration: The jagged frontier OpenAI measures how coding agents are accelerating AI research, where their performance remains uneven, and what this signals about recursive self-improvement. 🟢 AI How Sumitomo Life’s leaders helped teams adopt ChatGPT Enterprise Sumitomo Life uses ChatGPT Enterprise to research insurers, prepare proposals, and draft code for investment analysis. Early use by executives helped drive adoption. 🟢 AI The Silent Consensus Dean Ball, Head of Strategic Futures at OpenAI, examines how more capable AI could reshape society and how people can retain control of their institutions. 🟢 AI How OpenAI’s Performance Factory keeps ChatGPT fast OpenAI 推出性能工厂,利用 Codex 驱动的工作流查找和修复 ChatGPT 性能问题,保持其快速响应 🟢 AI Google Releases Gemini 4 Hydrogen: 2.5x Faster Than Gemini 3.8 Flash Google has released Gemini 4 Hydrogen , a fast model for coding and agentic work. Gemini 4 Hydrogen is a proprietary model from Google DeepMind built for everyday agent loops. It sits below the frontier models Gemini 4 Carbon and Gemini 4 Argon . Google calls it its “daily driver for agentic workflo 🟢 AI Google Releases TPU Sync: Open-Source KV Cache Transfer for TPUs Google Cloud has released TPU Sync , an open-source C++ library for moving KV caches and model weights on Cloud TPUs. The Google TPU Sync KV cache transfer library treats TPU HBM, host DRAM, shared storage and the data center network as one memory hierarchy. It targets a growing bottleneck: moving m 🟢 AI Perplexity Releases Decider v1.3: Open-Weight Decision Models Add Locate Perplexity has released pplx-decider-v1.3 and pplx-decider-flash-v1.3 , the latest versions of its open-weight decision models. Perplexity Decider v1.3 is a decision model that returns probabilities, not generated text. It answers yes/no, multiple-choice and score questions about a text or image inp 🟢 AI Best AI Video Generation Models in 2026: Elo, Price per Second, Open Weights The AI video leaderboard looks very different from 12 months ago. Native audio is now table stakes. Clips run 15–30 seconds. And an open-weights model sits in the top 3. This ranking (as of Nov 13, 2026) uses one quality signal: blind human-preference Elo from the Artificial Analysis Video Arena . W 🟢 AI Innovating with AI in the special chemicals industry As companies look to accelerate innovation with artificial intelligence, cloud computing, and advanced modeling, the primary challenge is connecting new technologies to reliable data, scientific expertise, and the practical realities of bringing products to market. For specialty chemicals manufactur 🟢 AI Nancy Pelosi’s replacement is ready to regulate AI A new voice for AI regulation in Congress: Scott Wiener, who authored the country's first proposed frontier AI legislation, just won Nancy Pelosi's former congressional seat and is already pushing for federal action — warning that "we could sleepwalk into a disaster." Recursive self-improvement is k 🟢 AI The backlash about data centers might really just be about electricity This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here . As voters went to the polls across the US last week, I was in Ashburn, Virginia, part of what’s known as Data Center Alley—the world’s largest data center hub. I 🟢 AI Nvidia privately confirmed it could track chips to fight smuggling, even as industry publicly questioned feasibility Representatives from Nvidia told Illinois representative Bill Foster that it would be technically possible to implement the anti-smuggling measures called for in the Chip Security Act, Foster told MIT Technology Review in an interview on Thursday, even while an industry group claimed they were “pote 🟢 AI Formally verifying agent-written full-stack web apps with PDD Kit At a glance Formal methods for everyday software. PDD Kit helps you turn your intent into a formal spec, then proves the code your agent writes for every input and uses tests where a proof can’t reach. It takes formal methods beyond niche, specialist work to engineers without a formal methods backgr 🟢 AI Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents Microsoft researchers have discovered “Lazarus,” a backdoor technique that can reprogram dormant neurons in an AI model, causing an otherwise normal AI agent to execute malicious commands when a hidden trigger appears. The threat is especially serious for AI agents with tool or terminal access. A tr 🟢 AI Microsoft Research Focus – Edition #98 .yt-consent-placeholder[data-video-id="Hcc6Ua7uTas"] { aspect-ratio: 9 / 16; max-width: 480px; margin: 24px auto 48px auto; } What’s new in research? In this edition of Microsoft Research Focus , based on the Research Focus newsletter (opens in new tab) , host Anson Ho shares some of the latest news 🟢 AI LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B 🟢 AI llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem 🟢 AI XYBench: Helping with the problem behind the question 🟢 AI Build a Robot Pick-and-Place Simulation with SimReady Assets 🟢 AI Introducing AlphaProof Code: a breakthrough in verifying complex software at depth AlphaProof Code agents wrote three million lines of proofs in Lean to verify Signal’s core encryption. 🟢 AI Gemini 4 Carbon is now available 🟢 AI AlphaGenome 2: Deciphering how cells read DNA to better understand disease Predicting how DNA sequences and their variations impact specific human cell types and disease. 🟢 AI Introducing Gemini 4 Hydrogen 🟢 AI Supporting research on AI and children’s safety OpenAI commits $2 million to research on AI and children’s safety. The work supports shared tools to measure risks and opportunities for young people. 🟢 AI Holiday Extras is rethinking what it can build with Codex With Codex, Holiday Extras is turning customer expertise into better apps, higher booking conversion and new possibilities for growth. 🟢 AI Gondola doubles revenue as its team builds more with Codex Gondola helps creatives get credit and get hired, with Codex raising its lead engineer’s output by 50% and ChatGPT simplifying daily work. 🟢 AI How Amplitude ships product improvements in 1–2 days with OpenAI Built primarily on OpenAI models, Amplitude’s proactive product agent, Wave, identifies what to build, helps teams ship changes with coding agents, and measures the impact. 🟢 AI Shield helps technicians hand off legacy portal work with OpenAI Shield uses OpenAI to help IT technicians hand off work in legacy portals. Its Forge agents navigate systems so technicians can focus on issues needing their judgment. 🟢 AI Introducing ChatGPT Wallet Introducing ChatGPT Wallet, a secure way for agents in ChatGPT Work, Codex, and dots to complete purchases with user approval and existing card protections. 🟢 AI Google just redesigned the search box for the first time in 25 years — here’s why it matters more than you think. For a quarter century, the Google search box has been one of the most recognizable interfaces in computing: a thin white rectangle, a blinking cursor, a few typed words, and a list of blue links. On Tuesday, Google will formally retire that paradigm. Read more 🟢 AI VentureBeat names Rob Strechay as its first Lead Analyst, expanding its enterprise AI research push Rob Strechay, until recently managing director and principal analyst at theCUBE Research, has joined VentureBeat as our first Lead Analyst and a founding analyst of VentureBeat Research. His arrival is the next step in a deliberate move at VentureBeat toward deeper specialization: analysis built for 🟢 AI Orchestration is the new challenge for CX in the age of AI agents Presented by Tata Communications Read more 🟢 AI When agents act on their own, governance has to live in the data layer Presented by EDB Read more 🟢 AI Enterprise AI's real risk isn't autonomous agents. It's the complexity between them. Presented by Gravitee Read more 🟢 AI Muse can now place calls to US businesses — if you pay Muse's AI can now place outbound calls to U.S. businesses — but it's not free. 🟢 AI Democrats launch Nvidia investigation, write scathing letter on Huang’s ‘scheme’ with Trump This Congressional investigation into Nvidia's China sales and its dealings with Trump is the second one, so far, since news of the alleged spy affair broke. 🟢 AI Jensen Huang didn’t show confidence in his successor, report says Huang described Kress as a "deer in the headlights" during a meeting in summer. 🟢 AI Israel’s VCs see opportunity for challengers in Nvidia’s turmoil Given Nvidia’s stronghold on AI and its importance in Israel, many in the country's startup ecosystem are worried about the impact of his departure. 🟢 AI ElevenLabs’ new app combines meeting note-taking and dictation for teams ElevenAI is a new meeting notetaker and dictation app aimed at teams and enterprises. 🟢 AI The unfolding Jensen Huang scandal: a timeline Jensen Huang has resigned from Nvidia following an affair, which could have ramifications for national security as the U.S. and China race to develop the most advanced AI infrastructure. 🟢 AI ChatGPT’s new Wallet lets AI agents shop and check out for you ChatGPT Wallet lets AI agents shop and check out on users' behalf, starting with a limited rollout to U.S. Pro subscribers. 🟢 AI Export-banned Nvidia chips are behind China’s most powerful AI models — uh-oh, Jensen! New research shows the most advanced AI models from China were trained on banned Nvidia chips. 🟢 AI AI agent startup LangChain hits $3B valuation LangChain now claims 7,000 customers of its model and platform agnostic agent management tool LangSmith. 🟢 AI Jensen Huang’s departure blasts big hole in AI startup funding, VCs and founders say Nvidia poured tens of billions into AI startups under Huang. Without him at the helm, founders and VCs are bracing for smaller checks and lower valuations. 🟢 AI 智谱 GLM 5.3 FlashX Legal:法律垂直领域专用模型 智谱 AI 11-02 发布法律领域专用版,定价低至 $0.007/M 输入。国内厂商正以垂直微调而非通用大模型的方式切入行业市场。 🟢 AI KT 发布 Mi:dm K 3.0 Pro 韩语大模型 韩国电信 KT 11-05 发布 Mi:dm K 3.0 Pro,主打韩语与本地合规场景,代表区域运营商自研模型路线仍在推进。 🟢 AI Google Gemini Omni 1.2 Flash:视频生成模型迭代 11-05 发布的视频生成模型,为 Omni 1.1 Flash 的次日迭代版本。Google 近 30 天已发布 6 个模型,为当前迭代最快的厂商。 🟢 AI Liquid AI 发布 LFM3-1.7B:小参数模型含免费档 1.7B 参数高效模型,11-06 上线并同步提供免费版本,适合边缘设备与本地推理。Liquid AI 近 30 天发布 3 个模型。 🟢 AI Meta 发布 Muse Ember 模型 Meta 11-09 发布 Muse Ember,为 Muse 系列新成员。Meta 近 30 天已发布 3 个模型,定价区间 $0.10/M 输入、1.0M 上下文。 🟢 AI Thinking Machines 发布 Inkling 2 Small 免费版 Mira Murati 创办的 Thinking Machines 11-09 发布 Inkling 2 Small 并提供免费档,是其小参数产品线的延续,可与开源权重方案直接比较。 🟢 AI Anthropic 推出 Claude Sheets、Whiteboards、Data 对标 Microsoft Office 在 IPO 前推出三款 AI 生产力工具,让用户用自然语言构建表格、白板与数据集。标志 Anthropic 从模型供应商转向办公套件竞争者。 🟢 AI Anthropic 推出 Claude Meetings:把会议讨论转成计划与任务 Claude Meetings 可解析工作场所讨论内容并直接产出计划与可执行任务,属 Anthropic 向办公协作场景延伸的一步,与 Claude Sheets、Whiteboards 构成组合。 🟢 AI 百度发布 ERNIE-5.5-Flash-Preview 与 ERNIE-Light-DT-1.0 百度 11-10 同日发布两款模型:ERNIE-5.5-Flash-Preview 主打低成本高吞吐,ERNIE-Light-DT-1.0 定位轻量端侧场景。国产模型保持双周级迭代节奏。 🟢 AI Gemini 4 Carbon 发布:Google 新一代旗舰 LLM Google 11-10 发布 Gemini 4 Carbon 及 batch 版本。在 Anthropic 公布的 AutomationBench 业务流程评测中领先,但 FrontierCode 2.0 得分 41.1% 低于 Fable 5.2 的 49.1%。 🟢 AI Fable 5.2 协调 1000+ 子 Agent 八天从零构建可启动操作系统 Anthropic 披露 Fable 5.2 从空目录起步,协调 1000+ 子 Agent 八天产出含内核、磁盘/网络驱动、GUI 与安装器的可启动系统。同时内部 90% 以上合入代码由 Claude 编写。 🟢 AI Anthropic 发布 Claude Fable 5.2 与 Mythos 5.2:$8/$40 每百万 token 两个模型同源不同权限:Fable 5.2 面向公众并限制敏感安全/生物任务,Mythos 5.2 仅对通过审核的机构开放。输出速度提升 30%,定价 $8/M 输入、$40/M 输出。 🟢 AI Reflection Beam 1.0: 501B 参数 MoE 模型 A23B 架构 Reflection 10月16日发布 Beam 1.0,总参数 501B、激活参数 23B 的 MoE 架构模型,定位大规模推理场景。 🟢 AI DeepSeek V4.1 Flash 发布:低成本推理新选择 DeepSeek 10月17日发布 V4.1 Flash,定价 $0.04/M tokens,上下文 1.0M。定位高性价比推理市场。 🟢 AI Moonshot AI 发布 Kimi K3.1 Pro 月之暗面 10月23日发布 Kimi K3.1 Pro,定价 $0.19/M tokens,上下文 1.0M。延续 Kimi 系列长上下文优势。