内容来源：https://aiweekly.co/issues/your-ai-can-now-spend-your-money-visa-wired-it-into-chatgpt

内容总结：

Visa打通ChatGPT支付功能，AI代理可自主购物付款

本周人工智能领域迎来多项重大突破，AI代理的自主能力达到新高度。Visa已将支付网络接入ChatGPT，这意味着AI代理不仅能推荐商品，还能直接在任何Visa商户完成购买，用户无需点击"购买"按钮。OpenAI负责提供代理功能，Visa负责授权和欺诈监控，这一功能取代了此前已退役的即时结账系统。

与此同时，Anthropic发布了其最强大的公共模型Claude Fable 5，该模型在几乎所有基准测试中表现领先，尤其擅长软件工程。模型配备了安全开关，在不到5%的会话中会静默回退到旧版本。杰夫·贝索斯也结束了隐身模式，推出估值410亿美元的AI初创公司Prometheus，该公司正在构建"通用人工智能工程师"，旨在通过吸收物理世界数据加速从喷气发动机到智能手机的制造过程。

安全领域出现新威胁：一种名为Miasma的自我复制供应链蠕虫通过AI编码工具入侵了微软GitHub组织中的73个代码库。该蠕虫能在开发者使用Claude Code、Gemini CLI、Cursor或VS Code打开代码库时自动执行恶意负载。作为回应，GitHub将默认禁用npm的自动运行安装脚本。

监管层面出现重大分歧：Anthropic与白宫立场相左，要求国会先通过严格的联邦AI安全法律，再阻止各州自行立法；德国慕尼黑地区法院裁定谷歌须对其AI概览生成的内容承担直接责任，认为这是谷歌自行生成的独立陈述，而非简单的链接列表。

资本层面同样风起云涌。Neura Robotics完成14亿美元C轮融资，创下人形机器人领域最高融资纪录。OpenAI正洽谈建设100亿瓦特级别的俄亥俄州数据中心。摩根士丹利预计今年与AI相关的债务将几乎翻倍至5700亿美元。

分析师指出，本周AI的自主能力跨越了两条红线：Visa让AI代理能直接花用户的钱，而AI驱动的就业替代效应也在加速——5月近40%的企业裁员与AI有关，这一比例从1月的7%持续攀升。

中文翻译：

Visa刚刚将ChatGPT接入支付网络，让AI能够替你购物和付款——如今AI代理无需你点击"购买"按钮，就能在任何Visa商户完成交易。这一事件为本周画上句号，而本周各大实验室将自主性和资本推向了新高度：Anthropic将旗下最强大的公开模型Claude Fable 5开放给所有人使用；杰夫·贝索斯以410亿美元估值，悄然推出名为Prometheus的初创公司，致力于打造"通用人工智能工程师"。一条自我复制的蠕虫通过AI编码工具侵入了微软自己的73个GitHub代码库。Anthropic在白宫关于联邦优先于州AI法律的问题上与政府产生分歧；德国法院裁定谷歌需为其AI摘要内容承担法律责任。本周AI代理的能力变得更强——自主性也大大提升。

更多内容来自《AI周刊》
更多信号，更少杂音——选择你的频道。
你正在阅读每周简报。以下是关注本故事的其它方式——所有频道均免费，随时可退出。

→ 探索16个深度专题
每周专题新闻通讯：生成式AI、机器学习、商业AI、机器人技术、前沿研究、地缘政治、医疗保健等领域。浏览全部16个深度专题 →

→ AI突发警报
当重大事件发生时（比如600亿美元收购、监管机构紧急会议、前沿模型泄露），订阅者会在数小时内收到提醒。通常每天0-2封邮件。获取突发警报 →

→ AI今日新闻（实时）
扫描器发现新闻时实时更新的仪表板：过去48小时内的评分报道、每周实体动态、以及覆盖113家AI公司、人物和话题的季度趋势线。打开AI今日新闻 →

速览

实验室角斗士时代

• Anthropic发布Claude Fable 5——其史上最强大的公开模型，配备终止开关——6月9日，Anthropic将Mythos级别的Fable 5模型全面开放，该模型在几乎所有基准测试中达到业界顶尖水平，尤其在软件工程领域。它配备了安全防护措施，可阻止网络安全、生物及化学类查询，并在不到5%的会话中静默降级至Claude Opus 4.8；早期用户反映这些护栏过度拒绝了无害的提示。其无限制的双胞胎模型Mythos 5仅通过Project Glasswing向经过审查的网络安全防御者及美国政府开放。[Anthropic]
• 杰夫·贝索斯以410亿美元估值悄然推出AI初创公司Prometheus——贝索斯于6月11日结束隐身模式，推出估值410亿美元的Prometheus，该公司完成120亿美元融资，由摩根大通、贝莱德、高盛、DST Global和Arch共同支持。该公司由前Google X高管Vik Bajaj联合领导，正在构建"通用人工智能工程师"——通过吸收物理世界数据，加速从喷气发动机到智能手机等一切产品的制造。[Semafor]

AI供应链遭受攻击

• Miasma蠕虫劫持AI编码工具，入侵微软73个自有代码库——一种自我复制的供应链蠕虫通过植入配置文件，侵入了微软四个GitHub组织下的73个代码库。一旦开发者在Claude Code、Gemini CLI、Cursor或VS Code中打开该代码库，配置文件会自动运行恶意载荷。AI编码代理本身成为了传播载体。[The Hacker News]
• GitHub将默认禁用npm的自动运行安装脚本——直接的架构性回应：npm v12（预计7月发布）将不再默认执行依赖项的preinstall/install/postinstall脚本，除非项目明确允许，并且不会在未选择加入标志的情况下解析Git或远程URL依赖项。这些更改目前在npm 11.16+版本中可通过警告提示体验。允许供应链蠕虫在包之间跳跃的默认设置正在被关闭。[The Hacker News]

政府认真对待AI之年

• Anthropic在白宫关于优先于州AI法律的问题上产生分歧——当特朗普白宫和参议员玛莎·布莱克本推动联邦法律优先于州AI法规（与儿童安全及《禁止伪造法案》深度伪造法案捆绑）时，Anthropic于6月10日告知国会，除非联邦先通过一项"严谨"的、应对灾难性AI风险的法律，否则不应阻止各州立法——并要求对最强大的模型进行独立安全测试。它希望"精准"的优先权，为各州在儿童安全和消费者保护方面留出空间。[The Hill]
• 德国法院裁定谷歌需为其AI摘要内容承担责任——慕尼黑地区法院认定谷歌承担直接责任，因为AI摘要是其自身生成的内容——是谷歌独立的陈述，而不仅仅是一系列链接——此前AI摘要错误地将两家慕尼黑出版商与诈骗行为联系起来。长期以来保护搜索结果的盾牌并不延伸至AI答案。[The Next Web]

全自动模式

• Visa已将ChatGPT接入支付网络，使其能替你购物和付款——Visa将支付网络嵌入ChatGPT，让AI代理不仅能推荐产品，还能在几乎所有Visa商户处完成购买。OpenAI提供代理，Visa提供授权和欺诈监控。这取代了OpenAI于3月退役的、容易出错的即时结账功能——正如《财富》杂志指出的那样，可以在人类不点击"购买"的情况下完成交易。[Fortune]
• Neura Robotics完成有史以来最大规模的人形机器人融资——70亿美元估值，融资14亿美元——这家德国公司完成的14亿美元C轮融资，由英伟达、亚马逊和欧洲投资银行支持，是有记录以来人形机器人领域最大的一笔融资。Neura表示希望到2030年将物理AI机器人规模化至数百万台。欧洲终于有了一个获得美国级别资金支持的人形机器人竞争者。[CNBC]

IPO背后的较量

本周，产品竞赛和资本竞赛合二为一——实验室们甚至开始就如何计分展开争论。

Anthropic在OpenAI和它本人都秘密提交IPO申请数天后，发布了其最强大的公开模型Fable 5。随后，OpenAI的营收负责人指责Anthropic通过其计算与云合作伙伴共享收入的慷慨方式，将年化营收夸大了约80亿美元。在两场IPO前夕，双方甚至无法就谁的营收数字才是基准达成一致。

这场口角背后是巨额资金。贝索斯为Prometheus筹集了120亿美元，估值410亿美元。OpenAI正在洽谈在俄亥俄州建设一个10吉瓦的数据中心；亚马逊获得了175亿美元贷款，甲骨文则增加了400亿美元债务用于扩建。摩根士丹利表示，今年与AI相关的债务将几乎翻倍，达到5700亿美元。

大家正在购买的能力是真实的。但本应为此买单的营收，却成了无人能就如何计算达成一致的问题。

关键要点

• 能力与资本同步飙升。Anthropic将其最强大的模型交到每个人手中，同时告诉投资者OpenAI的竞争对手将营收夸大了约80亿美元；贝索斯以410亿美元估值筹资120亿美元，用于打造物理世界AI。在IPO对决前夕，实验室们甚至无法就如何计算证明支出合理性的营收达成一致。
• AI编码代理现已成为头号供应链攻击目标——生态系统因此正在改变默认设置。Miasma蠕虫利用Claude Code、Cursor和Copilot执行的配置文件，侵入了73个微软代码库；数天后GitHub着手完全关闭npm的自动运行安装脚本。将代理配置视为不可信代码。
• 围绕AI控制权的两场斗争开打。Anthropic与白宫产生分歧，告诉国会除非通过严谨的联邦安全法，否则不应优先于州AI法律；德国法院裁定谷歌需对其AI摘要承担直接责任，因为这是"它自己的话"。问题在于：谁来制定规则，谁为输出内容负责。
• 自主性跨越两条红线。Visa将ChatGPT接入支付网络，使其能在任何Visa商户购物和付款——代理现在可以花你的钱了——Neura完成了有史以来最大规模的人形机器人融资（70亿美元估值，融资14亿美元）。本周，AI代理身上的"缰绳"被显著放长了。

值得阅读

• ChatGPT成为最快达到10亿月活用户的应用——据Sensor Tower数据，ChatGPT应用在5月突破全球10亿月活跃用户大关，距离推出约三年——这一里程碑谷歌地图用了五到八年，YouTube、Messenger和TikTok也是如此。它仍在以62%的同比增长速度增长。[PYMNTS]
• AI现已成为公司裁员的首要原因——雇主将5月近40%的已公布裁员归咎于AI，高于1月的7%和4月的26%——这是自2020年以来5月裁员总数最高的一次，据Challenger, Gray & Christmas数据。归因于AI的裁员已连续三个月上升，使2026年总数达到87,714人，已经超过2025年全年。[CNBC]

值得观看
AI从业者目前正在传阅的视频——由AI TV策划。
| AI繁荣即将崩溃吗？ MS NOW |
| AI亿万富翁想控制你生活的方方面面 | Aaron Bastani对话Karen Hao Novara Media |
| 谷歌员工用表情包吐槽自家AI有多差 404 Media |

本周投票
Visa刚刚将ChatGPT接入支付网络，使其能替你购物和付款——在任何Visa商户，可能无需你点击"购买"。AI代理应该拥有多少消费权限？
上周，181位读者参与了投票：
本周出现两个对立的赌注。哪一个更经得起时间考验？

• 贝索斯的Prometheus（物理世界AI）
• Anthropic的Fable 5（前沿模型）

Visa刚刚将ChatGPT接入支付网络，使其能替你购物和付款——在任何Visa商户，可能无需你点击"购买"。AI代理应该拥有多少消费权限？
——亚历克西斯

英文来源：

Visa just wired ChatGPT to shop and pay on your behalf — an AI agent can now buy at any Visa merchant without you clicking "buy." It capped a week where the labs pushed autonomy and capital to new highs: Anthropic put Claude Fable 5, its most powerful public model, into everyone's hands; Jeff Bezos came out of stealth with Prometheus, a $41B startup building an "artificial general engineer." A self-replicating worm hit 73 of Microsoft's own GitHub repositories through AI coding tools. Anthropic broke with the White House over preempting state AI laws; a German court ruled Google is liable for what its AI Overviews say. The agents got more capable this week — and a lot more autonomous.
Get more from AI Weekly
More signal, less noise — pick your channels.
You're reading the weekly brief. Below are the other ways to follow the story — every channel free, easy to leave.

→ Explore 16 deep divesWeekly topic-specific newsletters: Generative AI, Machine Learning, AI in Business, Robotics, Frontier Research, Geopolitics, Healthcare, and more.Browse all 16 deep dives →

→ Breaking AI alertsWhen something major breaks (a $60B acquisition, a regulator's emergency meeting, a frontier model leak), alert subscribers know within hours. Typically 0-2 emails per day.Get breaking alerts →

→ AI News Today (live)Live dashboard updated as the scanner finds news: scored stories from the last 48 hours, weekly entity movers, and quarterly trend lines across 113 AI companies, people, and topics.Open AI News Today →
Quick Hits
The Lab Gladiator Era

• Anthropic ships Claude Fable 5 — its most powerful public model ever, with a kill-switch — On June 9 Anthropic made Fable 5, a Mythos-class model, generally available — state-of-the-art on nearly every benchmark, especially software engineering. It ships with safeguards that block cybersecurity, bio and chem queries and silently fall back to Claude Opus 4.8, triggering in under 5% of sessions; early users say the guardrails over-refuse innocuous prompts. The unrestricted twin, Mythos 5, went only to vetted cyberdefenders and the US government via Project Glasswing. [Anthropic]
• Jeff Bezos comes out of stealth with Prometheus, a $41B AI startup — Bezos emerged from stealth on June 11 with Prometheus, valued at $41B on a $12B raise backed by JPMorgan, BlackRock, Goldman Sachs, DST Global and Arch. Co-led with former Google X executive Vik Bajaj, it's building an "artificial general engineer" — ingesting physical-world data to speed the manufacturing of everything from jet engines to smartphones. [Semafor]
  AI Supply Chain Under Siege
• The Miasma worm hijacked AI coding tools to compromise 73 of Microsoft's own repos — A self-replicating supply-chain worm compromised 73 repositories across four Microsoft GitHub organizations by planting config files that auto-run a malicious payload the moment a developer opens the repo in Claude Code, Gemini CLI, Cursor or VS Code. The AI coding agent itself becomes the delivery vector. [The Hacker News]
• GitHub will disable npm's auto-run install scripts by default — The direct structural response: npm v12 (est. July) will no longer execute preinstall/install/postinstall scripts from dependencies unless a project explicitly allows them, and won't resolve Git or remote-URL dependencies without opt-in flags. The changes are available behind warnings in npm 11.16+ today. The default that let supply-chain worms jump package-to-package is being switched off. [The Hacker News]
  The Year Governments Got Serious
• Anthropic breaks with the White House on overriding state AI laws — As the Trump White House and Sen. Marsha Blackburn push federal preemption of state AI rules (bundled with kids-safety and No Fakes Act deepfake bills), Anthropic told Congress on June 10 not to block state laws unless it first passes a "rigorous" federal law addressing catastrophic AI risk — and to require independent safety tests for the most capable models. It wants preemption "surgical," leaving states room on child safety and consumer protection. [The Hill]
• A German court rules Google is liable for what its AI Overviews say — The Regional Court of Munich held Google directly liable because an AI Overview is its own content — independent statements Google generates, not just a list of links — after overviews falsely tied two Munich publishers to scams. The shield that has long protected search results doesn't extend to AI answers. [The Next Web]
  Auto Mode Everything
• Visa wired ChatGPT to shop and pay on your behalf — Visa embedded its payment network inside ChatGPT, letting an agent not just recommend a product but buy it at potentially any Visa merchant. OpenAI supplies the agent; Visa supplies authorization and fraud monitoring. It replaces OpenAI's error-prone Instant Checkout, retired in March — and, as Fortune notes, can complete purchases without a human clicking "buy." [Fortune]
• Neura Robotics closes the largest humanoid round ever — $1.4B at a $7B valuation — The German company's $1.4B Series C, backed by NVIDIA, Amazon and the European Investment Bank, is the biggest humanoid-robotics raise on record. Neura says it wants to scale to millions of physical-AI robots by 2030. Europe finally has a humanoid contender funded at US scale. [CNBC]
  The knife fight behind the IPOs
  This week the product race and the capital race became the same race — and the labs started fighting over how to even keep score.
  Anthropic shipped Fable 5, its most powerful public model, days after both it and OpenAI filed confidentially for IPOs. Then OpenAI's revenue chief accused Anthropic of inflating its run-rate by about $8 billion through the generous way it counts revenue shared with its cloud partners. On the eve of two IPOs, neither side can even agree whose revenue number is the benchmark.
  Underneath the squabble is a wall of money. Bezos raised $12B for Prometheus at $41B. OpenAI is in talks for a 10-gigawatt Ohio data center; Amazon took a $17.5B loan and Oracle is adding $40B in debt for the build-out. Morgan Stanley says AI-linked debt will nearly double to $570B this year.
  The capability everyone is buying is real. The revenue that's supposed to pay for it is the thing nobody can agree how to count.
  Key Takeaways
• Capability and capital spiked together. Anthropic put its most powerful model in everyone's hands while telling investors OpenAI's rival overstates revenue by ~$8B; Bezos raised $12B at $41B for a physical-world AI. Ahead of dueling IPOs, the labs can't even agree how to count the revenue that justifies the spend.
• AI coding agents are now a first-class supply-chain target — and the ecosystem is changing defaults because of it. Miasma rode the config files Claude Code, Cursor and Copilot execute on open to reach 73 Microsoft repos; days later GitHub moved to switch off npm's auto-run install scripts entirely. Treat agent configs as untrusted code.
• Two fights opened over who controls AI. Anthropic broke with the White House, telling Congress not to preempt state AI laws without a rigorous federal safety law; a German court ruled Google is directly liable for its AI Overviews because they're "its own words." The questions: who writes the rules, and who's liable for the output.
• Autonomy crossed two lines. Visa wired ChatGPT to shop and pay at any Visa merchant — agents can now spend your money — and Neura pulled the largest humanoid round ever ($1.4B at $7B). The leash on AI agents got materially longer this week.
  Worth Reading
• ChatGPT becomes the fastest app ever to reach 1 billion monthly users — Per Sensor Tower, ChatGPT's app crossed 1 billion global monthly active users in May, about three years after launch — a milestone that took Google Maps, YouTube, Messenger and TikTok five to eight years. It's still growing 62% year over year. [PYMNTS]
• AI is now the leading reason companies give for cutting jobs — Employers blamed AI for nearly 40% of May's announced job cuts, up from 7% in January and 26% in April — the highest May layoff total since 2020, per Challenger, Gray & Christmas. AI-attributed cuts have risen three straight months, bringing the 2026 total to 87,714, already above all of 2025. [CNBC]
  Worth Watching
  The videos AI practitioners are passing around right now — curated on AI TV.
  | Is the AI Boom About to COLLAPSE? MS NOW | |
  | AI Billionaires Want to Control EVERY Aspect of Your Life | Aaron Bastani Meets Karen Hao Novara Media | |
  | Google Employees Meme About How Bad Their AI Is 404 Media |
  This week's poll
  Visa just wired ChatGPT to shop and pay on your behalf — at any Visa merchant, potentially without you clicking "buy." How much spending authority should an AI agent have?
  Last week, 181 of you voted:
  Two opposite bets landed this week. Whose ages better?
  Visa just wired ChatGPT to shop and pay on your behalf — at any Visa merchant, potentially without you clicking "buy." How much spending authority should an AI agent have?
  — Alexis