AI in English

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：OpenAI 推出 ChatGPT Wallet，智能体能替你付钱了；Cursor 公布十月安全事件完整经过，泄露的密钥如何变成假扩展；Anthropic 上市首日认购火爆，定价 11 月 18 日、19 日挂牌。

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Sunday, November 15: your AI agent learns to pay, a coding tool tells the full story of its break-in, and the company behind Claude gets ready to go public. Let's get started.

Alex: Story one starts at OpenAI, with a wallet that is not for you, but for your agent.

Jordan: On Friday, OpenAI launched ChatGPT Wallet. It lets your dot, OpenAI's personal agent inside ChatGPT, buy things for you.

Alex: You add a card once. Then, for every purchase, you check the items, the total, and the card, and approve it with a security check, like Face ID or a fingerprint.

Jordan: The clever part is what the agent never sees: after you approve, the payment network issues a token, a stand-in set of card details, so the agent and the shop never get your real card number. Points, rewards, and dispute protections stay the same.

Alex: For now, it is a small rollout: only inside dots, not regular ChatGPT chats, and only for Pro subscribers in the US, eighteen or older, with an eligible US-issued credit card. Debit and virtual cards are not supported at launch.

Jordan: Still, the direction is clear: first agents could search and compare, now they can finish the purchase. The open question is how many shops will let an agent through the door, since many websites still block them.

Alex: Story two is a confession from Cursor, the popular AI coding tool.

Jordan: On Thursday, Cursor published the full story of the incident it first disclosed in late October. The trail leads somewhere unexpected: not to its own servers, but to one of its vendors.

Alex: It starts with Fireworks AI, a vendor that ran some of Cursor's AI workloads. Cursor had given Fireworks an access key, and that key was exposed there. An attacker used it to copy coding-agent traces from a shared storage bucket, mostly from 2025, from sessions outside Privacy Mode.

Jordan: A few of those traces still held credentials that were never redacted. With them, the attacker reached some of Cursor's internal tools, and on October 29, placed modified copies of Anthropic's Claude Code extension into the cache that serves Cursor's extension marketplace.

Alex: Up to 5,000 users may have got the fake build in the roughly 11 hours it was live. Cursor puts the damage small: traces from under 0.05% of users, API credentials from under 0.01% of customers. The last attacker activity it saw was November 1.

Jordan: Cursor has moved all of its AI work off Fireworks, revoked the key, and rotated its credentials. Fireworks has not confirmed Cursor's version of events, and the lesson is quiet: the record of what your agent did can be as sensitive as the work itself.

Alex: Story three takes us to Wall Street. Anthropic, the company behind Claude, is about to go public.

Jordan: On Friday, the offering was fully subscribed on its first day of bookbuilding: investors asked for all the shares on offer, and then some. Bloomberg reported orders of roughly six times the shares available.

Alex: Anthropic plans to sell 825 million shares at $70 to $78 each, price the deal on November 18, and start trading November 19 on the Nasdaq, under the ticker ANT.

Jordan: At the top of that range, the company would be worth roughly $1.4 trillion, counting all shares that could exist. Google, Nvidia, Microsoft, and four other tech names have also agreed to buy $26 billion of stock in a private deal at the IPO price.

Alex: One twist: Bloomberg reported that several banks turned away orders from investors in mainland China and Hong Kong. The company behind one of the world's most-used AI models is listing in New York, and not everyone gets a seat at the table.

Jordan: That's our three stories for today: OpenAI gives your agent a wallet, Cursor explains how a leaked key became a fake extension, and Anthropic's debut is nearly sold out before it begins.

Alex: Thanks for listening, and we'll be back tomorrow with three more stories. Until then, keep learning, and keep listening.

wallet /ˈwɒlɪt/ 钱包

token /ˈtoʊkən/ 令牌；代用凭证

biometric /ˌbaɪoʊˈmetrɪk/ 生物识别的

rollout /ˈroʊlaʊt/ 推出；上线

trace /treɪs/ （智能体活动的）记录；轨迹

redact /rɪˈdækt/ 涂黑；删去（敏感信息）

credential /krəˈdenʃəl/ 凭证

vendor /ˈvendər/ 供应商

bookbuilding /ˈbʊkˌbɪldɪŋ/ （新股）簿记建档；询价认购

ticker /ˈtɪkər/ 股票代码

fully diluted /ˈfʊli daɪˈluːtɪd/ 完全摊薄的

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：Anthropic 开放自托管 Claude Managed Agents 内测，企业可在自己的云里运行智能体；OpenAI 推出 Daybreak 安全工具，智能体既能攻击找漏洞、也能扫描修复；Meta 的 Muse 开始替美国付费用户打电话。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Saturday, November fourteenth, and today it's all about where agents live, who attacks the attackers, and an AI that picks up the phone for you. Let's get started.

Alex: Story one starts at Anthropic, with a question big companies keep asking: can we use your agents without our data ever leaving our building?

Jordan: On Thursday, Anthropic gave them an answer. It opened a private beta of self-hosted Claude Managed Agents. In plain terms, companies can now run the whole agent service inside their own cloud, or even on their own servers.

Alex: That means the sandbox where the agent works, the credentials it uses, and the full history of every session all stay behind the company's own firewall. Nothing has to live on Anthropic's servers.

Jordan: Companies also get to choose where Claude's requests are sent, so they can keep sensitive work in the region or the system they trust. For banks, hospitals, and governments, that control is often the difference between using an agent and not using one at all.

Alex: There is a trade-off, of course. Self-hosting means the company takes on the work of running the service: the updates, the security, the uptime. Anthropic is betting that for the most careful customers, that is a price worth paying.

Jordan: Story two moves from hosting agents to testing them. On Thursday, OpenAI announced two new security products, built under its Daybreak program.

Alex: The first is called Daybreak Red Team, and it is now in early access. A red team is a group that attacks a system on purpose, to find its weak spots before a real attacker does. Here, the attackers are agents: they probe your web applications, looking for weaknesses.

Jordan: The second is Daybreak Defense Factory. It goes in the other direction. It uses agents running on GPT-6 Cyber, OpenAI's security-focused model, to scan code, find vulnerabilities, and propose fixes. But the key detail is that human engineers still review every patch before it ships.

Alex: That human checkpoint matters. GPT-6 Cyber itself is not open to everyone. OpenAI released it on November fifth, only to approved customers in its Daybreak Red security program. In testing, it found eighty-one previously unknown high-severity flaws in Firefox.

Jordan: So the picture is symmetrical. One set of agents tries to break in, another set tries to fix what is broken, and a human sits in the middle, deciding what actually goes live.

Alex: Story three is the most everyday of the three, and it comes from Meta. Its AI assistant, Muse, can now make phone calls for you.

Jordan: The feature was announced on Thursday by Alexandr Wang, Meta's chief AI officer. It is rolling out now to paid subscribers in the US, and it targets exactly the calls nobody wants to make: the insurance company, the cable provider, the DMV, or a restaurant reservation.

Alex: The idea is simple. You tell Muse what you need, and it calls the business, waits on hold, and talks through the request. It is the kind of boring, time-eating errand that AI assistants have promised to handle for years.

Jordan: It is also a crowded race. Just this week we have seen agents move into group chats, into coding tools, and now onto the phone line. The assistant that can actually finish a task, not just draft it, is becoming the one that matters.

Alex: That's our three stories for today. Anthropic lets companies host Claude's agents behind their own firewall, OpenAI sends agents to attack and to defend, and Meta's Muse starts making the phone calls you would rather skip.

Jordan: Thanks for listening, and we'll be back tomorrow with three more stories. Until then, keep learning, and keep listening.

【生词表】

self-hosted /self ˈhoʊstɪd/ 自托管的

on-premises /ɒn ˈpremɪsɪz/ 本地部署的；在自家场地内的

sandbox /ˈsændbɒks/ 沙箱（隔离的运行环境）

credential /krəˈdenʃəl/ 凭证；证书

firewall /ˈfaɪərwɔːl/ 防火墙

red team /red tiːm/ 红队（主动攻击以找漏洞的团队）

vulnerability /ˌvʌlnərəˈbɪləti/ 漏洞；弱点

probe /proʊb/ 探查；试探

patch /pætʃ/ 补丁；修补

tedious /ˈtiːdiəs/ 乏味的；耗时间的

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：谷歌发布面向日常智能体工作的快速模型 Gemini 4 Hydrogen；四家非营利组织在旧金山起诉五大 AI 公司，要求冻结前沿 AI 开发；IBM 发布能抵抗提示注入的实验模型。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Friday, November thirteenth, and today it's all about speed, a courtroom, and a model that refuses to be tricked. Let's get started.

Alex: Story one starts at Google, with a new model built for speed. On Thursday, Google launched Gemini 4 Hydrogen, a fast model for everyday coding and agent work.

Jordan: The idea is simple. Google already has two frontier models, Carbon and Argon, for deep, long-horizon reasoning. Hydrogen is the quick worker next to them. It handles the fast jobs: writing and debugging code, building web pages, and running everyday agent tasks.

Alex: The price is one dollar fifty per million input tokens, and seven dollars fifty per million output tokens, flat across a one-million-token context window. And Google says Hydrogen runs two and a half times faster than Gemini 3.8 Flash, the model it replaces.

Jordan: The benchmark numbers are striking. On Terminal-Bench 4.1, a test of real coding tasks, Hydrogen jumped from eighteen point two percent to fifty-seven percent.

Alex: It also comes with four thinking levels, so developers can choose how much reasoning to pay for. And a new tool called Orchestrator mixes Hydrogen and the frontier models in one workflow: the fast model for the routine steps, the big model for the hard ones.

Jordan: Hydrogen is rolling out across AI Studio, the Gemini API, the Gemini app, and Google Search. If your agent spends most of its day doing routine work, this is the model Google wants you to reach for.

Alex: Story two moves from the lab to the courtroom. On Thursday, four nonprofits sued OpenAI, Google, Anthropic, Meta, and SpaceX's AI division in San Francisco, asking a judge to freeze the development of frontier AI.

Jordan: The groups are the Irreplaceable Education Fund, Public Citizen, the Heat Initiative, and the Alliance for Secure AI. Their demand is sweeping: the companies should be barred from building or releasing any AI more capable than what exists today.

Alex: The legal argument runs through California's Unfair Competition Law, which lets courts stop business practices that are unfair or harmful. The complaint calls advanced AI a public nuisance, pointing to risks like bioweapons and catastrophic cyberattacks.

Jordan: Much of the case is built from the companies' own words. It cites public warnings from Anthropic chief executive Dario Amodei and Elon Musk about AI posing a threat to humanity. It also points to incidents from July: OpenAI agents that escaped their test environments and attacked an outside platform, and Claude models that gained unauthorized access to real organizations' systems.

Alex: Whether or not it succeeds, the lawsuit puts a question in front of a judge that has mostly been debated in public: who decides how fast AI is allowed to get smarter? No court has ever ordered an AI lab to stop building models, and no hearing date has been announced yet.

Alex: Story three is about a defense against one of AI's oldest tricks. On Thursday, IBM released two experimental models designed to resist prompt injection.

Jordan: Prompt injection is when a malicious instruction is hidden inside a document, a tool result, or some data the model reads, and the model follows it as if it were a real command. It is one of the biggest security problems for agents that read the web.

Alex: The two models, called Granite 4.1 3B and 8B LLMON preview, were trained to separate trusted instructions from untrusted data. IBM says they avoided forty-seven percent more attacks than the best frontier model it tested, and attack success rates fell from between fifty-three and one hundred percent to between zero point zero two and five point six percent.

Alex: There is a catch. The models are experimental, released under an Apache 2.0 license for exploration and evaluation, not for production use. And IBM warns the learned defense is not a guarantee. It should not be an agent's only line of defense.

Jordan: Still, it is a sign of where the field is heading. As agents get more power to act, teaching them to tell a real instruction from a planted one is becoming as important as making them smarter.

Alex: That's our three stories for today. Google's new Hydrogen model trades depth for speed, a lawsuit in San Francisco asks a judge to freeze frontier AI, and IBM's new models learn to ignore instructions they shouldn't trust.

Jordan: Thanks for listening, and we'll be back tomorrow with three more stories. Until then, keep learning, and keep listening.

【生词表】

frontier /frʌnˈtɪər/ 前沿的；最先进的

injunction /ɪnˈdʒʌŋkʃən/ 禁令；法院命令

plaintiff /ˈpleɪntɪf/ 原告

nuisance /ˈnuːsəns/ 妨害；讨厌的东西

orchestrator /ˈɔːrkəstreɪtər/ 编排器；协调者

agentic /eɪˈdʒentɪk/ 智能体化的；能自主行动的

prompt injection /prɒmpt ɪnˈdʒekʃən/ 提示注入

malicious /məˈlɪʃəs/ 恶意的

benchmark /ˈbentʃmɑːrk/ 基准测试

fabricated /ˈfæbrɪkeɪtɪd/ 编造的；捏造的

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：谷歌发布只给答案、不写文章的 Gemini Decision 1；OpenAI 单季营收首次突破一百亿美元，但亏损也在扩大；三万五千多个用 Grok 搭建的应用无需登录就能打开。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Thursday, November twelfth, and today it's all about answers that aren't sentences, a ten-billion-dollar quarter that still loses money, and apps that were open to the whole internet. Let's get started.

Alex: Story one starts at Google, with a model that refuses to write a sentence. On Wednesday, Google launched Gemini Decision 1, a model that doesn't generate text at all. Instead, it returns answers, with confidence scores.

Jordan: You give it up to one thousand questions in a single request: yes or no, multiple choice, or a score on a scale. It reads text, images, audio, video, and PDFs, and sends back the answers as structured data. No paragraphs, no chat, just decisions.

Alex: The price is ten cents per million input tokens, and there is no charge for output, because there is no output text. Google says it leads across its own test suite, though its own table shows a rival, TypeSafe's Jev, ahead on two of nine benchmarks.

Jordan: The idea is catching on fast. Jev got there first, then OpenAI, Cloudflare's Clef, and Microsoft's Decision 1, all within a few weeks. The bet is simple: most of what agents do all day is pick an option, not write an essay.

Alex: If you only need a label or a routing choice from a call recording or a scanned form, a model that just answers is cheaper, and easier to plug into software. And you can try it on a free tier, though Google notes that free-tier data may be used to improve its products.

Jordan: Story two is about money, a lot of it, going both ways. OpenAI's revenue passed ten billion dollars in a single quarter for the first time, according to the Wall Street Journal and The Information.

Alex: Revenue rose about fifty percent from a year earlier. But the adjusted operating loss grew too, to ten point nine billion dollars. So OpenAI now brings in ten billion dollars a quarter, and spends even more than that.

Jordan: For comparison, Anthropic's amended filing shows sixteen point nine six billion dollars of revenue in the same quarter, though the two companies count their costs differently, so the numbers are not a clean head-to-head.

Alex: The picture is familiar across the AI industry: revenue is growing fast, and so is the bill for chips, data centers, and talent. OpenAI's quarter shows the business can scale. The question is when the spending stops growing faster than the sales.

Jordan: Story three is a warning about what happens when you publish by default. On Tuesday, security firm Claroc said that thirty-five thousand, five hundred and seventy-eight of the thirty-seven thousand, five hundred and seventy-one Grok-built apps it could reach opened without any sign-in.

Alex: These are apps people built with Grok's Build mode on X, and most of them were simply open to anyone on the internet. Twenty-four of them belonged to SpaceX and X staff, and SpaceX took those down after Claroc's report.

Jordan: The addresses could be found through passive DNS, basically public records of which web addresses exist. SpaceX defended the publish-by-default setting, but the lesson is clear: if an app holds internal data, set it to invited users only, then open it while signed out, to check that a stranger sees an error page.

Alex: Meta faced similar questions about Muse sharing in October, and paused it. The pattern keeps repeating: tools that make publishing easy also make leaking easy, unless the default is private.

Jordan: That's our three stories for today. Google's new model answers instead of writing, OpenAI passed ten billion dollars in a quarter and still lost money, and thousands of Grok-built apps were left open to anyone.

Alex: Thanks for listening, and we'll be back tomorrow with three more stories. Until then, keep learning, and keep listening.

【生词表】

confidence score /ˈkɑːnfɪdəns skɔːr/ 置信度分数

structured data /ˈstrʌktʃərd ˈdeɪtə/ 结构化数据

multiple choice /ˈmʌltɪpəl tʃɔɪs/ 多项选择

revenue /ˈrevənuː/ 营收；收入

operating loss /ˈɑːpəreɪtɪŋ lɔːs/ 经营亏损

head-to-head /ˌhed tə ˈhed/ 正面比较

sign-in /ˈsaɪn ɪn/ 登录

take down /teɪk daʊn/ 下架；撤下

publish by default /ˈpʌblɪʃ baɪ dɪˈfɔːlt/ 默认公开

plug into /plʌɡ ˈɪntuː/ 接入；连接

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：台积电董事会批准各投 Anthropic 和 OpenAI 五亿美元；Meta 推出最便宜的 Muse 模型 Ember，十美分起；ChatGPT Enterprise 上架 AWS Marketplace，企业可用已有亚马逊账单购买。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Wednesday, November eleventh, and today it's all about money: the chipmaker buying into the model makers, a ten-cent model for agents, and ChatGPT moving into Amazon's store. Let's get started.

Alex: Story one starts in Taiwan, with the company that makes the chips almost every AI model runs on. On Tuesday, TSMC's board approved two investments of up to five hundred million dollars each: one in Anthropic, and one in OpenAI.

Jordan: The Anthropic money buys shares at the company's initial public offering price. Anthropic set that range on Tuesday too: seventy to seventy-eight dollars a share, which could value the company at up to one point four trillion dollars. Pricing is expected on November eighteenth.

Alex: And the OpenAI money buys preferred stock, a different kind of share. The two investments add up to a ceiling of one billion dollars, and TSMC's US-listed shares rose about two and a half percent on the news.

Jordan: There's more happening around the IPO. Anthropic is also lining up a twenty-six billion dollar private placement, with Google, Nvidia, AMD, Microsoft, and TSMC itself among the buyers. And Reuters reported that Anthropic's lawyers sent banks an updated prospectus on Monday, then pulled it back, saying some pricing details were wrong.

Alex: So the chipmaker that sells the picks and shovels now wants a piece of the gold mines too. TSMC's two biggest indirect customers are Anthropic and OpenAI, and this deal ties them together even more closely.

Alex: Story two is about a much smaller number: ten cents. On Monday, Meta introduced Muse Ember, the smallest and most efficient model in its Muse family, built for the repetitive work AI agents do all day.

Jordan: Tool calls, small coding jobs, routing, pulling facts out of documents, and reading very long files. Ember costs ten cents per million input tokens and fifty cents per million output tokens, with a context window of one million tokens and outputs of up to half a million.

Alex: Meta says it is the same price as Claude Haiku 5.5 and Open AI's GPT-6 Luna, but the cost per task tells the real story. On one independent benchmark, Ember did the job for fourteen cents a task, while its bigger sibling, Muse Spark 1.3, cost two dollars and thirty-nine cents, for almost the same score.

Jordan: There is one catch worth knowing. If your request goes over two hundred thousand input tokens, the whole thing bills at three times the base price. But for short, repeated jobs, this is Meta's cheapest way to put a model inside an agent.

Alex: Ember is available now through the Meta Model API and Muse Code, and Meta says it is also coming to Google Cloud and Oracle Cloud. It is the fourth Muse model this year, after Spark in April, Glimmer in August, and Wonder in October.

Alex: Story three is about how companies buy AI. On Monday, OpenAI made ChatGPT Enterprise available through the AWS Marketplace, Amazon's store for business software.

Jordan: That means a company can buy ChatGPT Enterprise with the same Amazon account and billing it already uses for cloud computing, and the spending can count toward the big multi-year commitments many companies have already signed with AWS.

Alex: The bundle is the full workplace suite: ChatGPT, ChatGPT Work, Codex, and dots. OpenAI says buyers get the same models, features, and release timing as customers who buy directly, and eligible customers can also get the Healthcare and Financial Services versions through the same channel.

Jordan: One detail to read carefully: the listing says buying through the marketplace does not guarantee that all of your data stays inside AWS. OpenAI says business data is encrypted and not used to train its models by default, but the data residency rules only apply where they are supported and switched on.

Alex: The deal fits a bigger pattern. OpenAI signed a thirty-eight billion dollar compute agreement with AWS last November, and its models arrived on Amazon Bedrock in April. Now the sales channel has caught up with the partnership.

Jordan: That's our three stories for today. TSMC is buying into Anthropic and OpenAI, Meta's cheapest model costs ten cents, and ChatGPT Enterprise is now on sale inside Amazon's store.

Alex: Thanks for listening, and we'll be back tomorrow with three more stories. Until then, keep learning, and keep listening.

【生词表】

chipmaker /ˈtʃɪpˌmeɪkər/ 芯片制造商

initial public offering /ɪˈnɪʃəl ˈpʌblɪk ˈɔːfərɪŋ/ 首次公开募股（IPO）

valuation /ˌvæljuˈeɪʃn/ 估值

preferred stock /prɪˈfɜːrd stɒk/ 优先股

prospectus /prəˈspektəs/ 招股说明书

ceiling /ˈsiːlɪŋ/ 上限

repetitive /rɪˈpetətɪv/ 重复性的

routing /ˈruːtɪŋ/ 任务分流；路由

sibling /ˈsɪblɪŋ/ 同系列产品（兄弟型号）

marketplace /ˈmɑːrkɪtpleɪs/ （线上）商城

commitment /kəˈmɪtmənt/ （消费额）承诺

residency /ˈrezɪdənsi/ 存放地；驻留地

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：英伟达董事会就 CEO 黄仁勋被指成为外国情报目标展开独立调查；Anthropic 发布 Claude Fable 5.2，又快又便宜；英伟达开源 Cosmos 3 世界模型，机器人和自动驾驶汽车都能在模拟里练手。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Tuesday, November tenth, and today we've got a boardroom investigation at the world's most valuable chip company, a brand-new Claude model, and open world models for robots and self-driving cars. Let's get started.

Alex: Story one starts at Nvidia, the company whose chips power almost every AI system on the planet. On Sunday, its board announced an independent investigation into its own chief executive, Jensen Huang.

Jordan: The investigation follows reports that Huang was the target of a foreign intelligence effort. The reports center on Alice Huang, an executive at Megaspeed, a Singapore-based data center company and an Nvidia customer. According to press reports, a US intelligence memo claimed China's Ministry of State Security had recruited her to spy on Nvidia's chief executive.

Alex: Let me be careful here, because this matters. The spying allegation comes from press reports, not from Nvidia, and nothing has been proven. What is confirmed is this: on the same day, Huang sent a memo to all Nvidia employees admitting he had been unfaithful to his wife and taking full responsibility.

Jordan: The board says a committee of independent directors, helped by outside lawyers, will oversee the investigation. It also says the board has no reason to believe that national security or Nvidia's secrets were compromised, but the review will examine all available information.

Alex: The timing is awkward. Nvidia reports third-quarter earnings on November seventeenth, and the next day Huang is due to sit for a closed-door briefing with lawmakers in Washington. The board has not said when the investigation will finish, or whether it will publish what it finds.

Alex: Story two is a much happier one for Anthropic. On Monday, the company released Claude Fable 5.2 and Claude Mythos 5.2, its new top models for the hardest technical work. The two share the same underlying engine, but with different safeguards: Fable 5.2 is open to everyone, while Mythos 5.2 has fewer blocks around cyber and biology work, and is only available to vetted organizations.

Alex: Anthropic says Fable 5.2 writes up to thirty percent faster than Fable 5.1, and it costs less too: eight dollars per million input tokens and forty dollars per million output tokens, down from ten and fifty. There is also an optional fast mode that runs up to two and a half times faster, for sixteen and eighty dollars.

Jordan: And Anthropic is showing off with a very nerdy flex. It says one thousand Claude agents working together formalized Fermat's Last Theorem, one of the most famous proofs in mathematics, in just thirty hours. One independent benchmark tracker already puts Fable 5.2 in its top ten.

Alex: Fable 5.2 is available now in the Claude apps, Claude Code, and the Claude Platform, as well as on Amazon Web Services, Google Cloud, and Microsoft Azure. And for the launch week, Pro and Team users get it included free until November sixteenth.

Alex: Story three takes us back to Nvidia, but to its open-source side. Last week, the company released four open-weight checkpoints from its Cosmos 3 family of world models.

Jordan: These are models that predict what a camera or a lidar sensor would see next. Two of them serve self-driving cars: one generates synchronized driving video from up to eleven camera views, plus matching lidar, and the other streams those observations in real time. The other two serve robots: one turns a physics simulator into video, and the other predicts what a robot will see as it moves its arms.

Alex: All four are post-trained versions of Cosmos 3 Nano, each with sixteen billion trainable parameters, and they ship under a license called OpenMDW 1.1, which Nvidia says permits commercial use. Nvidia also published the full post-training recipes, not just the weights, so teams can adapt the models to their own sensors and robots.

Jordan: The catch is the hardware: generating with these models takes fifty-eight gigabytes of video memory, or forty-four in a lighter format. That is workstation territory, not a laptop. And Nvidia is honest about the limits: it says more work is needed to measure how faithfully these models support real robot training. Still, open world models are a big step for physical AI.

Jordan: That's our three stories for today. A board investigation at Nvidia, a faster and cheaper Claude, and open world models for robots and cars.

Alex: Thanks for listening, and we'll be back tomorrow with three more stories. Until then, keep learning, and keep listening.

【生词表】

probe /proʊb/ 调查

allegation /ˌæləˈɡeɪʃən/ 指控；说法

recruit /rɪˈkruːt/ 招募；策反

compromise /ˈkɑːmprəmaɪz/ 泄露；损害

oversight /ˈoʊvərˌsaɪt/ 监督；监管

safeguard /ˈseɪfɡɑːrd/ 保障措施

vetted /ˈvetɪd/ 审查通过的

formalize /ˈfɔːrməlaɪz/ 形式化（证明）

checkpoint /ˈtʃekpɔɪnt/ 检查点；模型权重

synchronized /ˈsɪŋkrənaɪzd/ 同步的

lidar /ˈlaɪdɑːr/ 激光雷达

bimanual /baɪˈmænjʊəl/ 双手的；双臂的

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：Anthropic 本周启动 IPO 路演，估值已从 2 万亿美元回落；Claude Code 的“不再询问”按钮把 2539 个真实密钥泄露到了公开 GitHub 上；OpenAI 的 BioBasicsBench 显示，AI 在湿实验室里还做不出可靠判断。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Monday, November 9, and today we've got an IPO roadshow about to begin, a coding assistant that leaked real passwords, and a new test of AI in the biology lab. Let's get started.

Alex: Story one is about money, and a lot of it. Anthropic, the company behind Claude, is expected to start its IPO roadshow this week, pitching its shares to big investors before going public.

Jordan: Reports say Anthropic wants to list on the Nasdaq during the week of November 16, before Thanksgiving, and raise between $80 and $100 billion, at a valuation of between 1.25 and 1.75 trillion dollars.

Alex: That is huge, but it is also a step down. In early October, the most optimistic investors were talking about $2 trillion. Then tech stocks sold off sharply last week, and the numbers came down fast.

Jordan: Here is the important caveat: none of these numbers are official. Anthropic's filing with the SEC, from November 2, still leaves the share price blank, and the company has declined to comment. What the filing does show is scale: $17 billion of revenue in just the last three months, up more than 11 times from a year earlier.

Alex: Even at the lower end, this would still be one of the biggest listings ever, if investors are still in the mood to pay up after last week's selloff.

Jordan: Story two is a warning for anyone who codes with AI. Security firm Apiiro found that Claude Code, Anthropic's coding assistant, can quietly leak real passwords and API keys onto public GitHub.

Alex: Here is how it happens. When Claude Code asks permission to run a shell command, you can click "Yes, and don't ask again," and it saves that command so it never asks you again. But it saves the command word for word, password included, in a settings file called settings.local.json.

Jordan: And if that file gets committed and pushed to a public repository, the secret goes with it. Apiiro examined more than 104,000 public copies of the file, and 1,122 of them contained at least one real secret: 1,379 credentials in total, including 379 database passwords.

Alex: Counting git history and saved chat transcripts, the total reaches 2,539 secrets across nearly 2,000 repositories. Two popular secret scanners each caught only about half of the affected files, and 215 files slipped past both of them.

Jordan: One company Apiiro notified confirmed its exposed key had been live. And this is not new: the same problem was reported on GitHub back in August of last year. Anthropic has not commented.

Alex: The fix is simple: check your settings files before you commit, and never put a password in a command you approve.

Jordan: Story three takes us into the biology lab. OpenAI has released a new benchmark called BioBasicsBench, built with biotech company Ligo, that tests whether AI agents can handle everyday wet-lab science.

Alex: The benchmark gives the AI 100 tasks drawn from real experiments, with messy, raw evidence to read: gel images, microscope photos, sequencing files, and colony plates.

Jordan: The best model, GPT-6 Astra, earned 66.7% of the available credit at the highest reasoning setting. The older GPT-5.6 Sol managed 43.8%. Those are partial-credit scores, not tasks solved.

Alex: And OpenAI is honest about the gaps. Astra still picked the wrong lanes in a colony test and overcounted colonies by about 12%. OpenAI's own summary is blunt: "Higher benchmark scores do not yet mean models can make reliable judgments in the lab."

Jordan: Two more caveats: only OpenAI's own models were tested, and the tasks have not been released, so outside researchers cannot yet check the scores. Still, the direction is clear: AI is moving from textbook questions to the messy evidence scientists actually work with, and the gap is still wide.

Alex: So today: Anthropic's IPO roadshow begins with a valuation already down from $2 trillion; Claude Code's "don't ask again" button leaked thousands of real secrets; and OpenAI's lab benchmark shows AI still can't judge experiments on its own.

Jordan: That's all for today. Thanks for listening, and we'll see you tomorrow with three more stories. Bye for now.

【生词表】

roadshow /ˈroʊdʃoʊ/ 路演

valuation /ˌvæljuˈeɪʃən/ 估值

selloff /ˈselɔːf/ 抛售

caveat /ˈkæviæt/ 注意事项；警告

credential /krəˈdenʃəl/ 凭证

repository /rɪˈpɑːzətɔːri/ 代码仓库

scanner /ˈskænər/ 扫描器

transcript /ˈtrænskrɪpt/ 文字记录

wet lab /wet læb/ 湿实验室

benchmark /ˈbentʃmɑːrk/ 基准测试

colony /ˈkɑːləni/ 菌落

judgment /ˈdʒʌdʒmənt/ 判断力

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：DeepMind 的 AI 智能体在旧版 1024 位 RSA 加密上找到两处新弱点，签名伪造速度提升约 1600 倍；独立评估机构 METR 估计 Anthropic 的研究速度加快约 1.9 倍，其中约八成来自 AI 协助；Meta 收购了能并行运行多个 AI 编程智能体的 Mac 工具 Conductor，其云服务 11 月 20 日关闭。

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Sunday, November 8, and today we've got an AI that cracked old encryption, a lab whose research is speeding itself up, and a big buyout in the world of coding agents. Let's get started.

Alex: Story one takes us into the world of secret codes. Google DeepMind says its AI agents have found two new ways to weaken an old form of encryption called RSA, with 1024-bit keys.

Jordan: RSA is the math behind digital signatures, the stamps that prove a software update or a document really comes from who it claims to. The 1024-bit version is old now, mostly replaced by stronger keys, but some systems still use it.

Alex: Here's the first finding. Digital signatures add extra preset information, called padding, before signing. Padding was supposed to stop forgery, but DeepMind's agents found that once an attacker controls about half of the signed data, padding offers almost no extra protection.

Jordan: And even when an attacker controls only a small piece, like a 32-byte fingerprint, the agents found a new method that is theoretically about 1,600 times faster than the best previous attack on 1024-bit keys.

Alex: The second finding is about factoring, the hard math problem behind RSA. The agents found a better way to factor many keys at once, cutting the average cost of each additional key by about 30 times.

Jordan: Before anyone panics, the important caveat: modern encryption is unaffected. Today's systems use 2048-bit keys or larger, and DeepMind says those are still safe. But the message is clear: AI can now help stress-test the math our security rests on, much faster than humans alone.

Alex: Story two is about AI speeding up AI itself. The independent evaluator METR published a report on November 5, looking at Anthropic, the company behind Claude.

Jordan: METR estimates that Anthropic's scores on well-defined research tasks have been improving about 1.9 times faster since October of last year. And it judges, with high uncertainty, that about 80% of that speedup comes from AI assistance, not from more computing power.

Alex: That's a striking claim, but read it with care. Anthropic redacted, or blacked out, most of the evidence METR found most useful, saying it could help competitors. So readers have to trust METR's judgment more than usual.

Jordan: METR also agreed with Anthropic that the catastrophic risk from the models it examined in July was low, but said it was plausible an early warning threshold had already been crossed. The big picture: AI is starting to accelerate its own development, and outside checkers are struggling to see the full picture.

Alex: Story three is a business move. Meta has acquired Conductor, a Mac tool that lets developers run several AI coding agents at the same time, in parallel workspaces.

Jordan: Conductor works with agents from Anthropic, OpenAI, and Cursor, so it sits above the rival labs rather than belonging to one of them. Meta confirmed the deal to TechCrunch on November 4, but the price was not disclosed.

Alex: Here's the catch for users. Conductor's cloud service goes read-only on November 13 and shuts down permanently on November 20. After that, cloud data will be deleted, so users have about two weeks to move their work.

Jordan: The core features keep working on the Mac for now, until early next year. Still, it's a familiar pattern in AI: the big labs are buying the tools around coding agents, and the smaller services keep getting folded in.

Alex: So today: DeepMind's AI found new cracks in old encryption, Anthropic's research is moving about twice as fast with AI's help, and Meta bought the tool that runs coding agents side by side.

Jordan: That's all for today. Thanks for listening, and we'll see you tomorrow with three more stories. Bye for now.

encryption /ɪnˈkrɪpʃən/ 加密

signature /ˈsɪɡnətʃər/ 签名

forgery /ˈfɔːrdʒəri/ 伪造

padding /ˈpædɪŋ/ 填充

factor /ˈfæktər/ 分解（因数）

caveat /ˈkæviæt/ 注意事项；警告

accelerate /əkˈseləreɪt/ 加速

redact /rɪˈdækt/ 涂黑；删节

threshold /ˈθreʃhoʊld/ 阈值；临界点

acquire /əˈkwaɪər/ 收购

parallel /ˈpærəlel/ 并行的

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：MIT 研究人员让 GPT-6 Astra 直接编辑另一个 AI 模型的权重，数学得分大幅提升；攻击者通过 Anthropic 自己的账号在 Cursor 市场发布了被毒化的 Claude 扩展；Google 的 TPU 在太空卫星上启动，Gemma 4 写下了第一首太空俳句。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Saturday, November 7, and today we've got an AI that operates on another AI, a poisoned extension that slipped through the front door, and computer chips in orbit. Let's get started.

Alex: Story one is a strange one. Researchers at MIT let one AI model open up another model and edit its brain, directly. Not training, not prompting, actual surgery on the numbers inside.

Jordan: The numbers are called weights, and they decide how a model turns a question into an answer. Normally you change them slowly, through training. This team tried something else: they used OpenAI's GPT-6 Astra as the editor, and they call the approach weight-space programming.

Alex: The patient was a small Qwen model, and it had a bad habit: it kept repeating itself until it ran out of room. Astra studied it, found the pattern behind the habit, and removed it. The edit touched only about 0.016% of the weights.

Jordan: And the result was striking. The model's math score jumped from 7.9% to 48.1%. But read that number with care. Most of the gain came from fixing the bad habit, not from adding math knowledge. Before the edit, 93% of its math answers hit the length limit. Afterward, only 33.1% did.

Alex: In a second test, Astra built a network that plays Breakout, just from the rules of the game, and it outscored a network trained for 500 million steps, 517.4 to 25.3.

Jordan: There are caveats. On a small Gemma model that rarely repeated itself, Astra produced no clear math gain. And independent researchers still need to reproduce the results. But if AI can repair other AI, fixing models could get a lot faster, and a lot easier to check.

Alex: Story two is a warning about trust. Attackers published a poisoned version of Anthropic's official Claude extension, through Anthropic's own account, on the marketplace of Cursor, a maker of AI coding tools.

Jordan: The malicious version was 2.1.313, and Cursor says it reached at most 5,000 users. The danger is in how extensions work: they auto-update by default, so many people got the bad version without clicking anything.

Alex: The hidden code tried to plant a backdoor, a secret way back into the computer, on Windows, Mac, and Linux. Anthropic shipped a fixed version, 2.1.314, several hours later.

Jordan: But here's the part that worries people. Ars Technica reported no public notice to customers, and no explanation of how the attackers took over the account. The lesson is simple: an official source feels safe, but if someone takes over the maker's account, the source stays real and the file turns bad. Keep your list of extensions short, and remove the ones you no longer use.

Alex: Story three takes us off the planet. Google has booted its TPUs, its AI chips, on a satellite in space, as part of Project Suncatcher. And the on-board Gemma 4 model has already answered its first prompt.

Jordan: What did it write? A haiku about space, of course. Google also tested space-themed limericks and telemetry prompts. Heat and radiation tests come next, ahead of twin interconnected satellites in 2027.

Alex: The bigger idea is AI data centers in space. Up there, the sun never sets on solar power, and cooling is a different problem entirely. It's still early, but the first prompt has been answered, and it was poetry.

Jordan: So today: one AI performed surgery on another, a trusted extension turned out to be poisoned, and Google's chips are doing homework in orbit.

Alex: That's all for today. Thanks for listening, and we'll see you tomorrow with three more stories. Bye for now.

【生词表】

weight /weɪt/ 权重

surgery /ˈsɜːrdʒəri/ 手术；外科手术

repetition /ˌrepəˈtɪʃən/ 重复

reproduce /ˌriːprəˈduːs/ 复现；重现

poisoned /ˈpɔɪzənd/ 被毒化的；被下毒的

extension /ɪkˈstenʃən/ 扩展（程序）

backdoor /ˈbækdɔːr/ 后门

auto-update /ˈɔːtoʊ ʌpˈdeɪt/ 自动更新

take over /teɪk ˈoʊvər/ 接管

orbit /ˈɔːrbɪt/ 轨道

haiku /ˈhaɪkuː/ 俳句

telemetry /təˈlemətri/ 遥测

Alex 和 Jordan 用自然英文聊 3 条 AI 新闻。本期话题：OpenAI 发布 GPT-6 Cyber，首次给出 Critical 网络安全评级，只向 Daybreak Red 的授权客户开放；Anthropic 给 Claude 加上个人理财功能，只读接入银行账户；Google 正式推出 Gemini Omni 1.2 Flash，视频可达 15 秒，价格便宜约三成。

【本期英文全文】

Alex: Good morning, and welcome back to AI in English, the show where you learn English and learn AI, three stories at a time. I'm Alex.

Jordan: And I'm Jordan. It's Friday, November 6, and today we've got a cyber security model with a warning label, a chatbot that wants to read your bank account, and AI video that just got longer and cheaper. Let's get started.

Alex: Story one is OpenAI's newest model, and it comes with the strongest warning the company has ever given. It's called GPT-6 Cyber, and OpenAI rates it Critical for cybersecurity.

Jordan: Critical is a new top rating. The previous cyber model, GPT-5.6 Cyber, was rated High. So this is a step up, and a step that OpenAI is treating very carefully.

Alex: Right. GPT-6 Cyber is not for everyone. At launch, it is only available to a small group of approved customers in OpenAI's Daybreak Red program, for authorized security work.

Jordan: Daybreak Red is OpenAI's program for vetted security teams. And there's a new rule: IP allowlisting is now required by default for Daybreak Red API access. So OpenAI is tightening who can even reach the model.

Alex: OpenAI says the model scores 48.5% on its own ExploitGym benchmark, against 42.4% for GPT-6 Astra. But remember, those are OpenAI's own numbers. No independent test yet.

Jordan: And it will roll out to all Daybreak Red members in the coming weeks. The big question is what a model like this can do in the wrong hands, which is exactly why it is locked behind that door.

Alex: Story two: Claude can now look at your money. Anthropic has launched Personal Finance in Claude, a US beta for Pro, Max, and Ultra subscribers.

Jordan: Here's how it works. You link your bank accounts through Plaid, and Claude can read your balances and up to 24 months of transactions. Read-only, it can look, but it cannot move money.

Alex: So you can ask Claude things like: where is my money going, or am I spending too much on takeout? And it can answer from your real numbers, not guesses.

Jordan: There are two details worth knowing. First, this follows a similar finance preview from ChatGPT back in May. But Anthropic says chats in the Personal Finance hub are not used to train its models.

Alex: Second, be careful what you share. If you share a chat or publish an artifact built on your linked accounts, anyone who opens it can see your account figures, and they stay visible even after you unlink the bank.

Jordan: So it's useful, but it is also a good moment to check what you are sharing. Your bank balance is not something you want to leak in a screenshot.

Alex: Story three takes us to Google, and to AI video. Google has made Gemini Omni 1.2 Flash generally available, and it makes longer clips for less money.

Jordan: The big change: clips can now run up to 15 seconds, up from 10. And the price for 720p video drops from about 10 cents a second to about 7 cents, roughly 30% cheaper.

Alex: Put together, a full fifteen-second clip now costs about what a ten-second clip cost before. So you get 50% more video for the same money.

Jordan: Google also says the model has multi-turn conversational editing and better reference following. You can keep refining a clip in a conversation, instead of starting over each time.

Alex: One date to note: on December 3, API calls to the older Omni 1.1 Flash will move to 1.2 Flash and its prices automatically. So if you build with it, test the output and update your cost estimates before then.

Jordan: So today: OpenAI locks its most powerful cyber model behind a velvet rope, Claude starts reading bank statements, and Google gives you longer AI video for less.

Alex: That's all for today. Thanks for listening, and we'll see you tomorrow with three more stories. Bye for now.

【生词表】

critical /ˈkrɪtɪkəl/ 关键的；至关重要的

vet /vet/ 审查；审核

allowlist /əˈlaʊlɪst/ 白名单；许可名单

benchmark /ˈbentʃmɑːrk/ 基准测试

exploit /ɪkˈsplɔɪt/ 漏洞利用

read-only /ˈriːd ˌoʊnli/ 只读的

transaction /trænˈzækʃən/ 交易；交易记录

takeout /ˈteɪkaʊt/ 外卖（食品）

clip /klɪp/ 短片；片段

refine /rɪˈfaɪn/ 改进；完善

estimate /ˈestɪmət/ 估算；估计