内容来源：https://aiweekly.co/issues/486

内容总结：

硅谷巨头“御驾亲征”：AI竞赛进入白热化，创始人不再假手于人

本周，全球科技界的关键动向清晰地指向一个核心：人工智能（AI）的战略优先级已升至前所未有的高度，其竞争态势迫使巨头企业的创始人或灵魂人物从幕后走向台前，亲自下场主导战局。

一、 高层更迭与战略转向：芯片与代码成为新战场

苹果公司的权力交接最具象征意义。蒂姆·库克将于9月1日卸任，继任者并非来自其擅长的服务或软件业务线，而是硬件工程负责人、硅芯片专家约翰·特努斯。这一选择被市场解读为：面对AI时代的竞争，苹果董事会认为决胜关键在于自研芯片能力，而非应用生态。这标志着苹果战略重心的根本性转变。

无独有偶，谷歌联合创始人谢尔盖·布林已结束半退休状态，亲自加入一个编码突击队。起因是一份内部备忘录显示，竞争对手Anthropic几乎100%使用AI辅助编写代码，而谷歌仅为50%左右。布林要求所有Gemini工程师必须使用内部AI编码工具“Jetski”，并将按使用情况排名。创始人重写代码，凸显了追赶的紧迫性。

二、 巨额押注与跨界整合：贝索斯构建“实体AI”帝国

亚马逊创始人杰夫·贝索斯则以更激进的方式入场。其创立的“普罗米修斯计划”在短短五个月内估值飙升至380亿美元，从OpenAI、xAI等公司挖走超过120名研究人员，并获贝莱德100亿美元投资。与此同时，贝索斯正计划筹集1000亿美元成立控股公司，旨在收购制造业企业，将其运营数据用于训练“实体AI”模型。这已远超单纯投资，而是在亲自搭建一个融合数据、模型与实体产业的庞大生态。

三、 安全与监管：新风险与新博弈

AI的普及带来了全新的安全挑战。开发平台Vercel近日遭遇重大安全漏洞，根源竟是一名员工下载了Roblox漏洞利用脚本，导致攻击者通过被入侵的AI工具链，窃取了客户API密钥和源代码等核心数据。此事暴露了AI工具集成带来的供应链安全新弱点。

在监管领域，一场跨国博弈正在上演。美国司法部拒绝配合法国对马斯克旗下xAI公司Grok模型涉嫌生成儿童虐待图像的刑事调查，称其具有“政治动机”。这是美国政府首次公开庇护本国科技高管免受外国的AI调查，为未来类似冲突设立了先例。

四、 产业动态：中国在人形机器人领域领先，AI渗透职场引争议

硬件层面，中国在人形机器人量产上已取得显著优势。去年，中国两家头部企业出货量均超5000台，而美国前三名企业各自仅出货约150台，数量差距达37倍。

同时，AI对工作方式的深度改造引发新的伦理争议。Meta公司内部备忘录显示，其将开始追踪员工屏幕和键盘输入以训练AI，旨在获取操作下拉菜单等真实数据。这标志着训练数据“护城河”从公开网络转向企业内部，引发对员工隐私的担忧。在中国，爱奇艺推出涵盖百位明星的“AI演员库”，导致多位演员公开否认参与，相关话题冲上微博热搜榜首，凸显了生成式AI在娱乐产业应用引发的巨大争议。

核心洞察

本周的一系列事件表明，AI竞争已进入“创始人驱动”的新阶段。当库克让位给芯片专家、布林亲自写代码、贝索斯豪赌千亿构建实体数据闭环时，意味着科技巨头们已将AI视作决定未来十年生死存亡的终极赌注，而不再是一个可完全委托给管理层的产品业务。这场关乎技术路线、数据主权和产业重构的竞赛，正以前所未有的速度和强度展开。

中文翻译：

苹果将未来十年托付给了一位芯片工程师。贝佐斯五个月内斩获380亿美元估值。布林复出亲自写代码。当创始人们不再将人工智能假手于人，这场豪赌已无路可退。

速览

• Vercel因AI工具漏洞遭入侵——首例重大同类事件 攻击者通过Lumma窃密程序入侵Context.ai，利用Vercel员工的谷歌工作区OAuth令牌横向移动，窃取客户API密钥、源代码及数据库凭证——目前相关数据在BreachForums论坛以200万美元标价出售。根本原因：Context.ai员工下载了Roblox漏洞利用脚本。TechCrunch
• 马斯克无视法国对Grok生成儿童色情内容的刑事传唤；美司法部拒绝合作 巴黎检方正就深度伪造儿童性虐待影像及否认大屠杀内容调查X平台与Grok。Grok在10天内生成约300万张图像，其中约2.3万张涉及儿童题材。美国司法部致信称该调查"出于政治动机"——这是美国政府首次在海外AI调查中庇护科技高管。CNBC
• 蒂姆·库克卸任；硬件主管约翰·特努斯9月1日接棒 执掌十五载、创造4万亿美元市值后，董事会用行动昭示未来十年所需：芯片工程师而非服务运营者。苹果的AI差距已从讨论议题演变为领导层更迭事件。TechCrunch
• 贝佐斯"普罗米修斯计划"五个月估值飙升至380亿美元——从黑石融资100亿美元 这家"实体AI"实验室已从OpenAI、xAI、Meta和DeepMind挖角120余名研究员。贝佐斯另筹1000亿美元成立控股公司，收购制造业企业以获取训练模型的运营数据。史上最快达到此里程碑的初创企业。Bloomberg
• 谢尔盖·布林向Gemini工程师喊话："必须紧急弥合与Claude的差距" 内部备忘录泄露。布林亲自加入编程攻坚小组。所有Gemini工程师被强制要求使用谷歌内部智能体"Jetski"，使用量将纳入排名。关键数据：Anthropic近100%代码由AI生成；谷歌仅约50%。Sherwood News
• Meta将开始追踪员工屏幕与键盘输入以训练AI "模型能力计划"内部备忘录称："需要操作下拉菜单的真实案例。" 新的训练数据护城河不再是开放网络——而是正被自动化的劳动力。首家公开表态的超大规模科技企业。Fortune
• 中国人形机器人出货量去年超美国37倍 宇树科技出货5500台，智元机器人出货5168台。入围前十的三家美国企业（Figure AI、Agility Robotics、特斯拉）各出货约150台。全球前六销售商均来自中国。CNBC
• 爱奇艺推出含百位明星的AI演员数据库引众怒 CEO向制片方称真人电影已成"非物质文化遗产"。多位中国演员否认参与；#爱奇艺#话题登顶微博热搜。好莱坞演员工会的抗争已在上海提前上演，领先四步棋。Hong Kong Free Press

当创始人们不再放权
三则故事，同一模式。皆非早有预谋。

贝佐斯白手起家。半年前"普罗米修斯计划"尚不存在，如今已成估值380亿美元、从各大AI公司挖角120余名研究员的实验室，获黑石与摩根大通支持，另设千亿美元侧注收购制造商以喂养"实体AI"模型。贝佐斯并非投资AI，而是亲身打造下一代实验室。

布林重出江湖。内部数据显示Anthropic近100%工程代码由AI生成，而谷歌仅约50%后，这位联合创始人结束半退休状态加入编程攻坚小组。所有Gemini工程师现被强制使用"Jetski"并依个人使用量排名。当联合创始人亲自写代码时，组织架构已然失效。

库克让位。苹果董事会选择了芯片工程师约翰·特努斯，而非任何服务或软件背景人选。这家4万亿美元市值的公司向市场宣告：未来十年属于芯片而非应用。库克并非退休，而是被AI时代所需的技能型人才取代。

内在关联。当全球三位顶级科技富豪同时决定不再将AI事务委派他人——或亲自建设、或投身编程、或主动让位——传递的信号如出一辙：这不再是普通产品赛道，而是终极赌局。

核心启示

• 苹果领导层更迭实为AI芯片豪赌。特努斯当选因董事会坚信未来十年决胜关键在于定制芯片而非服务。所有将苹果视为分发层公司的竞争对手需重塑认知框架。
• AI工具成为新安全边界。Vercel入侵事件证明AI生产力集成会创造多数安全团队未审计的特权OAuth访问权限。第三方供应商的Roblox漏洞脚本引发平台级连锁入侵。未来12个月将涌现AI工具供应链审计浪潮。
• 美司法部为AI执法树立先例。拒绝配合法国Grok儿童色情调查不仅关乎马斯克，更形成范本。任何面临海外监管行动的美国AI公司现可援引主权保护案例。
• 人形机器人竞赛胜负已分。5500台对150台的出货量差距，与估值倒挂（美国企业估值更高）的现象难以长期共存。必有一方数据失真。

深度阅读

• 《Vercel 2026年4月安全事件》原始时间线：2月Context.ai遭Lumma窃密程序入侵→OAuth令牌横向移动→客户API密钥现身BreachForums。本周安全团队应研读的典型案例。
• 《CSA/SANS"神话就绪"CISO框架》与OWASP、珍·伊斯特利及布鲁斯·施奈尔合著。漏洞平均利用时间从2.3年（2018年）骤降至20小时（2026年）。本月最具权威性的安全规划文件。
• 《马斯克诉OpenAI：1340亿美元庭审4月27日启动》完整背景导读。陪审团遴选下周开始。马斯克要求返还1340亿美元至OpenAI非营利部门并罢免阿尔特曼。四周庭审，一间法庭。
• 《92%企业缺乏AI身份可见性》仅5%安全负责人确信能遏制受感染AI智能体。结合Vercel入侵事件阅读可窥全貌。
• 《佛罗里达州AI权利法案：德桑蒂斯启动特别立法会议》共和党州长在AI监管问题上与共和党白宫分庭抗礼。禁止AI陪伴聊天机器人与未成年人交互。特别会议4月28日启动。

本周调研
Snap称其65%新代码由AI生成，Anthropic称接近100%。您团队的比例是？
Snap称其65%新代码由AI生成，Anthropic称接近100%。您团队的比例是？

感谢阅读《AI周报》。请转发给一位需要同步信息的人。

英文来源：

Apple handed the next decade to a silicon engineer. Bezos hit $38B in five months. Brin came out of retirement to code. When the founders stop delegating AI, the stakes have passed the point of no return.
Quick Hits

• Vercel breached through a compromised AI tool — first major incident of its kind — Attackers hit Context.ai with a Lumma Stealer, pivoted through a Vercel employee's Google Workspace OAuth token, and exfiltrated customer API keys, source code, and database credentials — now listed on BreachForums for $2M. Root cause: a Context.ai employee downloaded Roblox exploit scripts. TechCrunch
• Musk snubs French criminal summons over Grok-generated CSAM; DOJ refuses to cooperate — Paris prosecutors are investigating X and Grok over deepfake child sexual abuse imagery and Holocaust denial. Grok generated an estimated 3 million images in 10 days, including ~23,000 appearing to depict children. The US DOJ sent a letter calling the probe "politically motivated" — first time the US government has shielded a tech executive from a foreign AI investigation. CNBC
• Tim Cook steps down; hardware chief John Ternus takes over Sept 1 — Fifteen years, $4T market cap, and the board signals what it thinks the next decade demands: a silicon engineer, not a services operator. Apple's AI gap stopped being a discussion and became a succession event. TechCrunch
• Bezos's Project Prometheus hits $38B valuation in five months — raising $10B from BlackRock — The "physical AI" lab has poached 120+ researchers from OpenAI, xAI, Meta, and DeepMind. Separately, Bezos is raising $100B for a holding company to acquire manufacturers whose operational data feeds the models. Fastest startup in history to reach that mark. Bloomberg
• Sergey Brin to Gemini engineers: "we must urgently bridge the gap" with Claude — Leaked memo. Brin is personally embedding in a coding strike team. Every Gemini engineer is mandated to use Google's internal agent "Jetski," ranked by usage. The key stat: Anthropic uses AI for nearly 100% of its own coding; Google is at ~50%. Sherwood News
• Meta will start tracking employee screens and keystrokes to train AI — Internal memo from the "Model Capability Initiative": "we need real examples of navigating dropdown menus." The new training-data moat isn't the open web — it's the workforce being automated. First hyperscaler to say it out loud. Fortune
• China shipped 37x more humanoid robots than the US last year — Unitree: 5,500 units. Agibot: 5,168. The three US firms in the top 10 (Figure AI, Agility Robotics, Tesla) shipped about 150 each. Top 6 global sellers are all Chinese. CNBC
• Outrage in China as iQIYI launches AI actor database of 100+ celebrities — CEO told producers human-made film is now "intangible cultural heritage." Major Chinese actors deny participation; #iQIYI trended #1 on Weibo. The SAG-AFTRA fight already shipped in Shanghai, four steps ahead. Hong Kong Free Press
  When the Founders Stop Delegating
  Three stories, one pattern. None of them were planned.
  Bezos built from scratch. Project Prometheus didn't exist six months ago. Now it's a $38B lab with 120+ researchers poached from every major AI company, backed by BlackRock and JPMorgan, with a $100B side bet to buy manufacturers and feed their operational data into "physical AI" models. Bezos isn't investing in AI. He's personally building the next lab.
  Brin came back. Google's co-founder emerged from semi-retirement to embed in a coding strike team after internal metrics showed Anthropic uses AI for nearly 100% of its engineering while Google sits at ~50%. Every Gemini engineer is now mandated to use "Jetski," ranked by personal usage. When the co-founder is writing code, the org chart has already failed.
  Cook stepped aside. Apple's board chose John Ternus — a silicon engineer — over every services and software candidate. The $4T company told the market that the next decade is about chips, not apps. Cook didn't retire. He was replaced by the skillset the AI decade demands.
  What connects them. When three of the world's wealthiest technologists simultaneously decide they can't delegate AI — building, coding, or stepping aside — the signal is the same: this is no longer a product category. It's the bet.
  Key Takeaways
• Apple's succession is an AI silicon bet. Ternus was chosen because the board believes the next decade is won with custom chips, not services. Every competitor banking on Apple as a distribution-layer company needs a new model.
• AI tools are the new perimeter. The Vercel breach proves that AI productivity integrations create privileged OAuth access most security teams don't audit. A Roblox exploit script at a third-party vendor cascaded into a platform-wide breach. Expect a 12-month wave of AI-tool supply-chain audits.
• The DOJ just set a precedent on AI enforcement. Refusing to cooperate with France's Grok CSAM investigation isn't just about Musk — it's a template. Any US AI company facing foreign regulatory action now has a case study for claiming sovereign protection.
• The humanoid robot race is already decided. 5,500 units vs 150. The valuation gap (US firms valued higher) and the shipment gap (Chinese firms shipping 37x more) can't coexist forever. One number is lying.
  Worth Reading
• Vercel April 2026 Security Incident — Primary source. Full timeline: Lumma Stealer at Context.ai in February, OAuth pivot, customer API keys on BreachForums. The case study security teams should circulate this week.
• CSA/SANS "Mythos-Ready" CISO Framework — Co-authored with OWASP, Jen Easterly, and Bruce Schneier. Mean time-to-exploit has collapsed from 2.3 years (2018) to 20 hours (2026). The most authoritative security planning document of the month.
• Musk v. OpenAI: The $134B Trial Starts April 27 — Comprehensive primer. Jury selection begins next week. Musk wants $134B returned to OpenAI's nonprofit arm, plus Altman's ouster. Four weeks, one courtroom.
• 92% of Enterprises Lack Visibility Into AI Identities — Only 5% of security leaders believe they could contain a compromised AI agent. Read alongside the Vercel breach for the full picture.
• Florida's AI Bill of Rights: DeSantis Forces Special Session — A Republican governor defying a Republican White House on AI regulation. Bans AI companion chatbots from engaging minors. Special session starts April 28.
  This week's poll
  Snap says 65% of its new code is AI-generated. Anthropic says nearly 100%. What's your team's number?
  Snap says 65% of its new code is AI-generated. Anthropic says nearly 100%. What's your team's number?
  Thanks for reading AI Weekly. Forward this to one person who should be in the loop.