🤖 AI 日报 | 今日十大热点（2026.06.05）

来源：Curious Spark · 每日 AI 精选

1. 🏆 美国五大AI巨头全面签署CAISI预发布安全审查协议：AI监管进入制度化里程碑 — NIST / The Hill / The Next Web / Bloomberg

据美国商务部下属NIST披露，截至2026年6月5日，微软、谷歌DeepMind和xAI已正式加入CAISI（人工智能标准与创新中心）预发布审查计划，与此前已加入的OpenAI和Anthropic合并构成美国五大前沿AI实验室全覆盖。各实验室须在新模型公开发布前向政府团队提供访问权限，由国家安全专家审查网络安全漏洞、自主代理行为及其他高危能力。白宫正评估将该自愿框架升级为行政命令，标志着AI监管进入制度化实质推进阶段。

NIST disclosed that Microsoft, Google DeepMind, and xAI have formally joined the CAISI pre-release review program, joining OpenAI and Anthropic to give full coverage of all five major U.S. frontier AI labs. Each lab must grant government teams pre-deployment access to test for cybersecurity risks, agentic behaviors, and other high-risk capabilities. The White House is evaluating upgrading this voluntary framework to an executive order, marking a landmark step toward institutionalized AI regulation.

2. 🤖 英国CMA首创全球「AI内容退出权」：强制谷歌允许出版商拒绝AI训练与搜索使用 — TechCrunch / CMA / ghacks.net

英国竞争与市场管理局（CMA）以「战略市场地位」为由，发布针对谷歌的历史性命令：所有在英国运营的出版商可通过Google Search Console一键退出AI Overviews、AI Mode等生成式AI搜索功能，且不影响自然搜索排名。出版商同时可拒绝其内容被用于训练谷歌旗下独立AI模型，并将获取内容在AI产品中使用的详细数据报告，谷歌须在AI摘要中明确标注内容来源。该裁定被视为全球「AI内容退出权」的开创性先例。

The UK CMA issued a landmark order against Google under its 'strategic market status' designation: all UK publishers can opt out of generative AI search features—including AI Overviews and AI Mode—via Google Search Console without affecting organic rankings. Publishers can also refuse their content from standalone AI model training and will receive detailed usage reports. Google must provide clear attribution in AI summaries, establishing a groundbreaking global precedent for 'AI content opt-out rights.'

3. 💡 Meta AI商务助手覆盖全球百万商户：WhatsApp/Instagram自动完成预约销售，B端AI落地爆发 — Meta / AI Weekly / The Next Web

Meta正式宣布其AI商务助理已覆盖全球超百万家企业，支持WhatsApp、Instagram和Messenger三大平台原生集成。商家无需额外开发即可接入AI助理自动处理客户预约、销售咨询、订单跟进等核心业务流。该功能将进一步整合Meta广告平台，AI对话数据用于优化广告精准度，助力中小企业以极低成本实现24小时智能客服和销售自动化，B端AI落地进入大规模扩张阶段。

Meta officially announced its AI Business Agent has reached over one million businesses globally, with native integration across WhatsApp, Instagram, and Messenger. Businesses can automate appointments, sales inquiries, and order follow-ups with zero custom development. The feature further integrates with Meta's ad platform, using conversation data to enhance ad precision, enabling SMBs to achieve 24/7 intelligent customer service at minimal cost as B2B AI deployment enters mass expansion.

4. 🚀 中国Robotaxi驶入全商业闭环：北上深全无人出租车规模盈利，L4智驾里程碑到来 — IDC / AITOP100 / AIBase / 知乎

据行业报告，无人驾驶出租车已在北京、上海、深圳实现全商业闭环运营，三地合计在运全无人驾驶车辆超数千辆，均在无安全员状态下于开放道路提供服务。主要运营商包括百度萝卜快跑、文远知行等，多个运营区域已实现单车盈利。2026年中国Robotaxi成为全球首个真正商业化的L4级自动驾驶赛道，预计年底前运营规模将扩大至五大城市，无人驾驶出租车商业元年正式到来。

Industry reports confirm fully autonomous robotaxis have achieved commercial viability in Beijing, Shanghai, and Shenzhen, with thousands of driverless vehicles serving the public without safety drivers. Major operators including Baidu's Apollo Go and WeRide have achieved per-vehicle profitability in multiple zones. China's robotaxi sector becomes the world's first truly commercialized L4 autonomous driving track in 2026, with expansion to five cities expected by year-end—the official commercial dawn of driverless taxis.

5. 🎯 比亚迪「尧舜禹」人形机器人年内量产两万台：汽车供应链颠覆机器人成本曲线 — AIBase / AIHub / 36Kr / 2026年中国AI与机器人报告

比亚迪旗下人形机器人项目「尧舜禹」进入量产冲刺，计划2026年内在自有工厂部署两万台，成为中国人形机器人量产规模最大的单一厂商。「尧舜禹」行走速度达1.5米/秒，手部配备触觉传感器，在焊接、物料搬运等工序效率已达人类工人80%。比亚迪将汽车供应链管理经验和零部件通用化策略引入机器人制造，大幅压低单台成本，打破「人形机器人必然昂贵」的行业定式，被视为中国制造业AI化的标志性里程碑。

BYD's humanoid robot project 'Yao Shun Yu' is entering mass production, with plans to deploy 20,000 units in its own factories within 2026, making it China's largest single-company humanoid robot manufacturer. The robot walks at 1.5m/s with tactile-sensing hands, achieving 80% of human worker efficiency in welding and material handling. BYD applied automotive supply chain and parts standardization strategies to robotics, dramatically lowering per-unit costs and breaking the 'humanoid robots must be expensive' paradigm—a landmark milestone in China's manufacturing AI adoption.

6. 📡 腾讯联手手机厂商推A2A跨平台AI助手：系统级跨App调度直达微信，移动AI深度融合 — AIBase / AIHub / AI快讯

腾讯联合多家安卓手机厂商正式推出A2A（App-to-App）系统级AI助手，实现通过一句指令跨应用自主完成任务，无需用户手动切换App。最具代表性的能力是AI助手可直接调起并操控微信，完成发消息、截图分享、预约提醒等原本需要手动操作的流程。该项目被视为对苹果Apple Intelligence的直接竞争，同时抢先布局腾讯即将推出的微信AI Agent生态，将系统级AI能力与国内最大社交平台深度绑定，移动AI时代进入新纪元。

Tencent and major Android manufacturers launched an A2A (App-to-App) system-level AI assistant enabling multi-app task completion with a single instruction—no manual app switching needed. Most notably, the AI can directly invoke and control WeChat for messaging, screenshot sharing, and reminders. Seen as a direct challenge to Apple Intelligence, the initiative stakes out early ground in Tencent's upcoming WeChat AI Agent ecosystem, deeply binding system-level AI with China's largest social platform.

7. 🧠 Uber人力部门AI预算半年耗尽：AI工具大规模替代HR岗，全球科技巨头AI替岗冲击波提速 — AIBase / AI Weekly / AITOP100

据内部知情人士透露，Uber人力资源部门2026年全年AI工具预算已在上半年全部耗尽，原因是AI驱动的招聘筛选、合规审查和员工绩效评估工具使用量远超预期，已替代了大量原本由HR专员处理的日常工作。Uber被迫在年中追加AI预算，并同步将人力岗位缩减约20%。该案例被视为AI自动化「从生产线蔓延至白领工作」的代表性样本，在全球互联网和企业服务领域引发对AI替岗速度超预期的广泛关注。

Internal sources reveal Uber's HR department exhausted its entire 2026 annual AI tools budget in the first half of the year, driven by far-higher-than-expected adoption of AI-powered recruiting, compliance review, and performance management tools that replaced large amounts of HR specialist work. Uber was forced to approve mid-year supplements while reducing headcount by ~20%. This case has become a representative example of AI automation spreading from factory floors to white-collar work, raising concerns that AI job replacement is outpacing projections globally.

8. 💰 IDC报告：2026年中国具身智能市场破110亿美元，年增速120%，制造服务农业全面渗透 — IDC / 2026年中国机器人与具身智能市场报告

IDC最新报告显示，2026年中国具身智能机器人市场支出规模将超110亿美元，同比增速达120%，成为全球增速最快的AI硬件细分市场。制造业是最大应用场景，占据约45%市场份额，其次是物流仓储（28%）和商业服务（17%）。IDC预计2027年中国具身智能市场将超越北美跃居全球第一，硬件价格下降和国产大模型能力提升是驱动增长的双核心引擎，AI物理世界时代加速到来。

IDC's latest report reveals China's embodied AI robot market will surpass $11 billion in 2026, growing at 120% YoY, making it the world's fastest-growing AI hardware subsegment. Manufacturing is the largest sector at ~45% share, followed by logistics (28%) and commercial services (17%). IDC projects China will surpass North America by 2027 to lead globally, driven by declining hardware costs and improving domestic LLM capabilities, as the physical AI era accelerates.

9. 🔬 中兴腾讯联手发布AI云电脑「Work Buddy」：自然语言驱动全套办公流程，企业AI桌面时代到来 — AIBase / AI快讯 / AITOP100

中兴通讯与腾讯联合发布AI云电脑产品线，搭载「Work Buddy」企业智能体，用户可通过自然语言一站式完成撰写报告、安排日程、数据汇总分析等全套办公流程，无需切换不同软件。该平台深度集成腾讯会议、企业微信和腾讯文档，支持跨部门协作工作流智能编排。双方计划年底前将AI云电脑覆盖至百万台企业终端，与微软Copilot+PC路线正面竞争，中国智慧办公市场正式进入AI原生工作台时代。

ZTE and Tencent jointly launched an AI Cloud PC product line featuring the 'Work Buddy' enterprise AI agent, enabling end-to-end office workflows—drafting reports, scheduling, and data analysis—through natural language without app switching. The platform deeply integrates Tencent Meeting, WeCom, and Tencent Docs. Both companies plan to cover one million enterprise endpoints by year-end, directly competing with Microsoft's Copilot+ PC roadmap as China's smart office market enters the AI-native workstation era.

10. 🌐 大模型安全攻防实测：GPT-5.5漏洞检测最强，DeepSeek V4 Pro成本最低，Gemini最难被越狱 — AIBase / 安全研究 / 2026年大模型安全报告

安全研究机构对最新一代主流大模型进行实战网络攻防测试后公布结果：GPT-5.5在已知漏洞利用检测和代码注入分析上表现最优，但API调用成本较高；DeepSeek V4 Pro以最低成本完成多数攻防任务，性价比最优；Gemini由于内置严格拒绝机制在大多数攻击测试中主动拒绝执行，被认为最难被「越狱」；Claude 4.6在多轮上下文威胁场景识别中表现突出。各模型安全差异化表现为企业AI安全基础设施选型提供重要参考。

Security researchers published real-world offensive security test results across major LLMs: GPT-5.5 achieved the best overall performance in vulnerability detection and code injection analysis but at higher cost; DeepSeek V4 Pro completed most attack-defense tasks at lowest cost; Gemini's strict built-in refusal mechanisms made it hardest to jailbreak; Claude 4.6 excelled at multi-turn contextual threat recognition. The differentiated security profiles provide important guidance for enterprise AI security infrastructure decisions.

📌 关注 Curious Spark，每日追踪 AI 前沿动态
如有转载请注明来源