Scroller Briefings -OpenAI Officially Launches Codex Plugins, Helping Millions of Developers Package and Share Configurations with One Click 🚀
编辑
目录
单词学习：![[2026032869785024.mp3]]
新闻播报：![[2026032807854222.mp3]]
OpenAI正式上线Codex插件功能，助力百万开发者实现配置一键打包与共享
简报：
- 3月27日，OpenAI正式为Codex推出插件（Plugins）服务，支持开发者将Skills（技能）、应用集成及MCP（模型上下文协议）服务器配置打包为可安装的工作流集合。
- 插件专为组织和团队协作设计，支持内部或市场分享，团队成员一键安装即可瞬间同步统一的开发配置，免去手动调优的繁琐。
- Codex明确划分了工具定位：技能（Skills）侧重个人定制及单一代码库迭代且无法发布；而插件（Plugins）则用于团队成员复用相同的集成方案。
- 插件生态市场已初具规模，首批接入了Slack、Figma、Notion、Cloudflare等头部生产力工具，未来将开放开发者将自研插件加入官方目录的权限。
- 受益于GPT-5.2-Codex的发布及功能更新，Codex平台整体使用量实现翻倍，上月开发者用户规模已正式突破100万。
相关链接：
OpenAI Officially Launches Codex Plugins, Helping Millions of Developers Package and Share Configurations with One Click 🚀
Brief:
- On March 27, OpenAI officially launched Plugins for Codex, enabling developers to package Skills, app integrations, and MCP (Model Context Protocol) server configurations into installable workflow bundles.
- Plugins are specifically designed for organizations and team collaboration, supporting both internal and marketplace sharing. Team members can instantly sync a standardized development setup with one-click installation, eliminating the hassle of manual configuration and tuning.
- Codex clearly distinguishes the roles of its tools: Skills are focused on personal customization and iteration within a single codebase, and cannot be published; Plugins, by contrast, are intended for team members to reuse the same integration setups.
- The plugin ecosystem marketplace has already begun to take shape, with the first batch of integrations including leading productivity tools such as Slack, Figma, Notion, and Cloudflare. In the future, developers will be allowed to add their own custom-built plugins to the official directory. 🔧
- Driven by the release of GPT-5.2-Codex and related feature updates, overall usage of the Codex platform has doubled, and the number of developer users officially surpassed 1 million last month. 📈
Related link:
谷歌全球推出Search Live功能，搭载Gemini 3.1开启实时多模态交互
简报：
- 谷歌宣布在全球超200个国家和地区正式上线“Search Live”功能，支持Android和iOS用户通过手机摄像头及语音指令与现实环境进行实时AI对话。
- 该功能由全新的原生多语言音视频大模型Gemini 3.1 Flash Live驱动，具有轻量化和高响应特性，显著提升了对话自然度与速度。
- 用户只需将摄像头对准物体（如复杂家具组装或动植物识别）并提问，系统即可同步提供语音解答及相关网页链接。
- 此举是谷歌应对Luma AI（Uni-1模型）及OpenAI等AI搜索竞争压力的关键战略，旨在巩固移动端入口壁垒，推动AI助手向“主动感知伙伴”进化。
相关链接：
Google Launches Search Live Globally, Powered by Gemini 3.1 for Real-Time Multimodal Interaction 🌍📱
Briefing:
- Google has officially launched the "Search Live" feature across more than 200 countries and regions worldwide, enabling Android and iOS users to engage in real-time AI conversations with their surroundings through smartphone cameras and voice commands.
- The feature is powered by the new native multilingual audio-video large model, Gemini 3.1 Flash Live, which is designed to be lightweight and highly responsive, significantly improving conversational naturalness and speed.
- Users simply need to point their camera at an object—such as during complex furniture assembly or plant and animal identification—and ask a question. The system can then provide spoken answers along with relevant web links simultaneously.
- This move is a key strategic response by Google to mounting competition in AI search from players such as Luma AI (with its Uni-1 model) and OpenAI. It is intended to strengthen Google’s mobile entry-point advantage and accelerate the evolution of AI assistants into “proactive perceptive companions.” 🤖
Related Link:
Anthropic新一代AI模型Claude Mythos进入内测，性能全面超越Opus
简报：
- 据3月27日曝光的Anthropic内部泄露文件显示，新一代超强模型Claude Mythos已进入秘密测试阶段。
- 文件披露了全新的模型层级命名“Capybara”（内部信息显示其与Mythos极可能指向同一底层架构），该层级规模更大、智能水平更高，综合能力已全面超过此前的旗舰模型Claude Opus。
- Anthropic内部风险评估指出，该模型带来了前所未有的网络安全风险，这也促使公司在追求智能与安全平衡时保持谨慎的发布节奏。
- 此次新模型的动向表明，下一代大模型正从单纯的对话能力向更深层次的逻辑推理与复杂任务处理演进。
相关链接：
Anthropic’s Next-Generation AI Model Claude Mythos Enters Internal Testing, Outperforming Opus Across the Board 🚀
Briefing:
- According to leaked internal Anthropic documents revealed on March 27, the next-generation high-performance model Claude Mythos has entered a confidential testing phase.
- The documents also disclose a new model tier named “Capybara” (internal information suggests it is highly likely to point to the same underlying architecture as Mythos). This tier is larger in scale, more intelligent, and has already surpassed the previous flagship model Claude Opus in overall capabilities.
- Anthropic’s internal risk assessment indicates that the model introduces unprecedented cybersecurity risks, prompting the company to maintain a cautious release pace as it seeks to balance intelligence and safety.
- The latest developments surrounding this new model suggest that the next generation of large models is evolving beyond simple conversational ability toward deeper logical reasoning and more advanced handling of complex tasks. 🤖
Related Link:
苹果计划在iOS 27开放Siri核心，全面接入第三方AI助手
简报：
- 知名科技记者于当地时间3月26日透露，苹果计划在即将发布的iOS 27操作系统中，向外部人工智能助手开放Siri的底层核心。
- 苹果正在开发全新的集成工具，未来用户从App Store安装的第三方AI聊天机器人（例如Claude等）能够与Siri实现深度整合。
- iOS 27用户将可通过单一指令，直接在Siri界面内无缝调用不同厂商的AI能力，无需在多个应用程序之间反复切换，实现自由切换不同的“AI大脑”。
- 苹果此举旨在确保Siri作为用户进入AI世界第一入口的中心地位，并借助外部成熟技术补齐Siri在复杂逻辑推理和多轮对话等方面的技术短板。
相关链接：
Apple Plans to Open Siri’s Core in iOS 27, Fully Integrating Third-Party AI Assistants 🤖🍎
Brief:
- A well-known tech journalist revealed on March 26 local time that Apple plans to open Siri’s underlying core to external AI assistants in the upcoming iOS 27 operating system.
- Apple is developing a brand-new integration framework that will allow third-party AI chatbots installed from the App Store in the future, such as Claude, to achieve deep integration with Siri.
- iOS 27 users will be able to use a single command to seamlessly access AI capabilities from different providers directly within the Siri interface, without repeatedly switching between multiple apps, enabling them to freely switch between different “AI brains.”
- This move is intended to ensure Siri’s central role as users’ primary gateway into the AI world, while also leveraging mature external technologies to address Siri’s current shortcomings in areas such as complex logical reasoning and multi-turn conversations. ✨
Related link:
腾讯正式发布Agent产品全景图，集成多款主流大模型并打通微信等国民级入口
简报：
- 腾讯正式对外发布 Agent 产品全景图，首次系统性展示其在智能体领域的顶层设计与底层架构（被称为“养虾车间”）。
- 系统底层由 AI Infra 智算底座提供算力，中间层 TokenHub 平台集成自研混元大模型，并兼容 MiniMax、百川智能、智谱 GLM、DeepSeek 及 Kimi 等主流模型。
- 配备丰富的 Skills 工具箱，除了集成腾讯文档、腾讯会议等自研原生工具，还通过外部 Skillhub 深度打通微信、企业微信、QQ 及元宝等入口，实现从“信息交互”到“任务执行”的转变。
- 构建了贯穿始终的全流程安全体系，结合 Lighthouse 原生安全、AI Agent 安全中心、腾讯 iOA 和电脑管家 18.0，有效管控数据隐私合规，防范 AI 误操作风险。
相关链接：
Tencent Officially Unveils Its Agent Product Panorama, Integrating Multiple Mainstream Large Models and Connecting to Mass-Market Entry Points Such as WeChat 🚀
Briefing:
- Tencent has officially released its Agent product panorama to the public, systematically showcasing for the first time its top-level design and underlying architecture in the intelligent agent space (internally referred to as the “shrimp farming workshop”).
- At the system foundation, the AI Infra intelligent computing base provides computing power, while the middle-layer TokenHub platform integrates Tencent’s self-developed Hunyuan large model and is also compatible with mainstream models such as MiniMax, Baichuan Intelligence, Zhipu GLM, DeepSeek, and Kimi.
- It comes equipped with a rich Skills toolbox. In addition to integrating Tencent’s native in-house tools such as Tencent Docs and Tencent Meeting, it also deeply connects external access points including WeChat, WeCom, QQ, and Yuanbao through Skillhub, enabling a shift from “information interaction” to “task execution” 🤖
- Tencent has also built an end-to-end security framework spanning the entire workflow. Combined with Lighthouse native security, the AI Agent Security Center, Tencent iOA, and Tencent PC Manager 18.0, the system effectively manages data privacy and compliance while preventing risks caused by AI misoperations 🔒
Related link:
主流AI应用隐私收集加剧：平均采集14项数据，70%获取位置信息
简报：
- 网络安全公司Surfshark报告显示，主流AI应用平均收集14种用户数据，目前高达70%的AI应用会获取用户位置信息，相比2025年的40%大幅上升。
- Meta AI收集数据多达33种（占比约95%），是唯一获取财务信息的AI应用，并收集种族、性取向等极度敏感信息；谷歌Gemini收集23种数据，含精准位置与设备通讯录。
- ChatGPT数据收集量同比暴涨70%（从10种增至17种），新增粗略位置、健康与健身等数据；Claude收集13种数据，部分内容被用于营销与分析。
- 鉴于AI机器人常处理税务、医疗等易外泄的高机密文件，安全专家建议用户将所有AI对话视为公开信息，并主动关闭聊天记录功能。
相关链接：
Mainstream AI Apps Intensify Privacy Collection: Averaging 14 Data Types, 70% Access Location Information 🔒📍
Brief:
- A report from cybersecurity company Surfshark shows that mainstream AI apps collect an average of 14 types of user data. Currently, as many as 70% of AI apps access users’ location information, a sharp increase from 40% in 2025.
- Meta AI collects as many as 33 types of data (about 95%), making it the only AI app that gathers financial information. It also collects highly sensitive data such as race and sexual orientation. Google Gemini collects 23 types of data, including precise location and device contacts.
- ChatGPT’s data collection volume surged 70% year over year, rising from 10 types to 17. Newly added categories include coarse location, health, and fitness data. Claude collects 13 types of data, some of which are used for marketing and analytics.
- Given that AI chatbots often handle highly confidential documents prone to leakage, such as tax and medical records, security experts advise users to treat all AI conversations as public information and proactively disable chat history features. ⚠️
Related link:
如果对你有用的话，可以打赏哦
打赏
本文作者:topwind
本文链接:
版权声明:本博客所有文章除特别声明外，均采用 BY-NC-SA 许可协议。转载请注明出处！
目录