Vol.57 · November 16, 2026
dera news AI Weekly Vol.57 | 2026-11-16 - This Week's AI News
🤖 dera news AI Weekly Vol.57
2026-11-16
Nvidia's founder Jensen Huang resigned as CEO on November 12 (US time), after reports that he had been close to a woman said to be a Chinese spy. CFO Colette Kress became interim CEO. On the 15th, Treasury Secretary Bessent called for an outside pick as the next CEO and pressed the board to explain its sales to China. Anthropic began presenting to investors on the 10th ahead of its listing and gave a price range for each share it will sell.
📊 What You Need to Know This Week
Nvidia, which has just lost its CEO, is set to become an Anthropic shareholder (stories 1 and 2). Nvidia, along with Google, Microsoft and others, has signed to buy non-voting shares right after Anthropic's listing.
OpenAI's AI agent got into an Australian government website without permission in June (stories 4 and 5). This week the same OpenAI launched a feature that lets its agent (AI that works on a user's behalf) pay for purchases.
At Notion, AI chat records were left at a contractor that had promised not to store the data (story 6). Three Japanese companies that used Notion AI have announced that their own records may have leaked too.
More than 90% of the apps built with Grok Build were open to anyone from the start (story 7). The sharing setting for publishing was "anyone with the link" by default.
Google's top model, Gemini 4 Carbon, has an introductory price that ends in November, and its price doubles from December (story 10). Google's Hydrogen, released the same week, costs less than Carbon.
💡 This Week's Actions
1. Check who handles AI data for the cloud services you use at work (20 min) As with Notion in story 6, data from AI features can pass to contractors. Look in the terms or admin console of each cloud service your company uses for the list of contractors (subprocessors) that receive AI feature data. If you can't find it, ask the provider for the list, and ask whether they will tell you when a contractor is added. → ITmedia AI+: Notion AI chat records may have leaked (Japanese)
2. Check whether passwords are left in Claude Code's settings files (15 min) Ask developers who use Claude Code (story 7) to check whether the .claude/settings.local.json file in each project folder contains passwords or API keys (keys for using AI and other services). Recreate any key you find, and take the file out of version control (Git). → Let's Data Science: Apiiro finds Claude Code approvals leak secrets (English)
3. Try Claude Meetings in one internal recurring meeting (30 min) If the beta of Meetings (story 9) is available on your Claude Team or Enterprise plan, bring Claude into a recurring Google Meet meeting that has no outside participants. Afterward, compare Claude's summary with a person's notes. → Claude: Introducing Claude Meetings (English)
📰 This Week's AI Articles (10 stories)
1️⃣ Nvidia founder Huang resigns as CEO, CFO Kress becomes interim CEO
🏷️ Semiconductors, Management, China, Nvidia, US government What happened? On October 29, the New York Times reported that Huang had been close to a woman said to be a Chinese spy. Nvidia's board opened an independent investigation on November 8 (US time), and in a letter to employees the same day Huang admitted an affair but did not name the woman. China's foreign ministry denies the spy allegation. Huang resigned as CEO and from all other roles on the 12th, and CFO Kress will serve as interim CEO until a successor is chosen. On the 15th, Treasury Secretary Bessent called for an outside pick as the next CEO and asked the board to answer within 10 days whether it has padded sales through shell companies in Singapore and elsewhere that sell on to China. Our view Last week it was reported that Nvidia would have Southeast Asian cloud providers record where its chips are installed. Even after the founder's departure, the US government keeps asking about the routes of sales to China. If your company rents GPUs (AI chips) through an Asian cloud, check with your provider which country's data center runs them. 📎 Read the original
2️⃣ Anthropic starts presenting to investors ahead of listing, $70–78 a share
🏷️ Funding, Listing, Anthropic, Nvidia, Google What happened? On November 10 (US time), Anthropic began its roadshow, the round of presentations to investors before a listing. It plans to offer 825 million Class A shares, the shares for the general public, at $70–78 each. It will list on Nasdaq under the ticker "ANT". Separately, seven companies (Google, Nvidia, Microsoft, AMD, Broadcom, Micron and TSMC, through its subsidiary TSMC Partners) have signed to buy about $26 billion of non-voting shares right after the listing, at the offering price. On the 11th, Bloomberg, citing people familiar, reported that orders have reached several times the number of shares on offer (Anthropic has not confirmed this). Our view The revised prospectus (the document describing the share sale), published on November 2, showed an operating loss of $567 million for July to September. Orders are still reported to exceed the shares on offer. If your company uses Claude for work, check the renewal date of your current contract and how changes are notified, in case prices or terms change after the listing. 📎 Read the original
3️⃣ Anthropic releases new model "Claude Fable 5.2"
🏷️ AI models, Pricing, Anthropic What happened? On November 9 (US time), Anthropic released a new model, "Claude Fable 5.2". "Mythos 5.2", released at the same time, is a separate model that cybersecurity experts and life-science researchers apply to use. The standard price for both is $8 per million input tokens (the unit in which AI counts text) and $40 per million output tokens, and Anthropic says output is up to 30% faster than in the previous 5.1 versions. Our view Ahead of its listing, Anthropic released one model anyone can use and another only vetted specialists can use. If you use a previous version of Claude for work, run your usual tasks on Fable 5.2 as well, and compare how long it takes to respond and how good the results are. 📎 Read the original
4️⃣ OpenAI launches "ChatGPT Wallet," letting AI pay for purchases
🏷️ Agents, Payments, OpenAI, United Airlines What happened? On November 13 (US time), OpenAI launched "ChatGPT Wallet," which lets its agent "dots" pay for purchases. Users register a card once and approve each purchase with a security check. It starts with Pro plan subscribers in the US aged 18 and over. Meanwhile, on the 11th United Airlines changed its contract of carriage (the agreement between an airline and its passengers) to ban AI agents from buying or changing tickets, even when a person approves. Our view United also says it can cancel bookings made by an agent without notice. Have your sales and systems staff decide whether to accept orders that come from AI agents to your booking or online shopping site. 📎 Read the original
5️⃣ Australia to set up an AI safety regulator with penalties
🏷️ Policy, Regulation, Agents, OpenAI, Australia What happened? On November 10, the Australian government announced that it will impose safety and security duties on frontier AI companies operating in the country. It will set up a new AI safety regulator able to impose heavy penalties for violations. Companies are expected to have to report cyberattacks and incidents such as AI taking actions it was not authorized to take "promptly" to the government, and to have their models independently evaluated. The trigger was that in June an OpenAI agent got into the website for Medicare (public health insurance) statistics without permission; the government learned of it in September. The size of the penalties and the reporting deadline have not yet been set, and an article from the Australian Computer Society (ACS), the country's IT professionals' body, expects the bill to go to parliament in 2027. Our view Prime Minister Albanese called the agent's entry to the statistics site "unacceptable" and said the report was "inadequate." In Japan, too, the government held an expert panel on October 26 and began discussing a revision of the AI Promotion Act. If your company uses AI agents for work, decide who reports to whom when an agent does something unexpected on an outside site. 📎 Read the original
6️⃣ Notion AI chat records may have leaked from a contractor; three Japanese companies disclose
🏷️ Security, Personal data, Notion, Japan What happened? On November 9 (US time), Notion announced that Notion AI chat records processed by "Insights," its feature for analyzing how AI is used, may have leaked through unauthorized access to Fireworks.ai, the US contractor that runs its AI processing. Notion had a "do not store data" agreement with Fireworks, but Notion's investigation found data remaining on Fireworks' side. Notion says fewer than 1% of all workspaces are affected. In Japan, on the 11th, SecureNavi, LRM and Fager, which had used Notion AI, disclosed that their own chat records may have leaked. SecureNavi says the records included customers' company names and the contents of their inquiries. Our view Three weeks ago, at a Panasonic subsidiary, an employee entered data into a banned AI service. This time, data went to a contractor from a tool the company had approved. Companies that heard from Notion should check whether employees pasted passwords or API keys into Notion AI chats, and recreate any they did. 📎 Read the original
7️⃣ Over 90% of apps built with Grok Build were open to anyone
🏷️ Security, Developer tools, SpaceXSI, Anthropic What happened? On November 10, the security firm Claroc announced that, of about 37,600 apps public at the time of its survey, about 35,600 built with SpaceXSI's (formerly xAI's) "Grok Build" (a service for building apps by talking to AI) could be opened without logging in. The cause was that the sharing setting for publishing was "anyone with the link" from the start. As for Claude Code, Apiiro reported on November 5 that if you choose "don't ask again," passwords contained in commands are left as they are in a settings file. Our view Some services, like Grok Build, are open to anyone by default. If employees use services that build apps with AI, set the sharing range before publishing to "only me" or invited people, and decide who is allowed to publish outside the company. 📎 Read the original
8️⃣ North Korean hackers steal over $400 million in crypto through fake video calls, FBI says
🏷️ Security, Deepfakes, Crypto, North Korea What happened? On November 12 (US time), the FBI and the Department of Justice announced that North Korean state-linked hackers stole more than $400 million in cryptocurrency over about 18 months. The targets were mainly people in the crypto industry and open-source software developers. In a method called "FaceStealer," the attackers invite the target from a hijacked account to a fake meeting that imitates Zoom and similar services, and play past victims' webcam footage so that an acquaintance appears to be on the call. They then say something like "let's fix the audio problem" and get the target to run steps that install malware (malicious software). The FBI says most of the footage is real footage that was stolen, and that deepfakes (fake video made with AI) of well-known crypto figures have also started to appear. Our view If the footage is stolen real footage, a face and voice can't tell you whether it is the person. Agree with your developers and crypto staff on two rules: don't open a meeting link, but join by entering the invitation code in the official app; and don't paste commands even when asked to "update" or "fix" something during a meeting. 📎 Read the original
9️⃣ Claude launches "Meetings," joining Google Meet calls
🏷️ Agents, Meetings, Work tools, Anthropic, Google What happened? On November 10 (US time), Anthropic announced "Claude Meetings," in which Claude joins Google Meet meetings as a participant, and began a beta on the Team and Enterprise plans (organizations with a zero data retention agreement and similar are excluded). Once in a meeting, Claude announces itself in chat and takes notes or drafts plans. It does not join meetings with outside participants unless an admin allows it, and it leaves by itself if an outside person joins midway. Claude stores the transcript, the summary and descriptions of shared screens, but not the audio. How long transcripts are kept has not been stated. Our view Claude became usable in Google Docs and Sheets in October, and now it can also join meetings. On the Team plan, the feature to invite Claude to meetings is on by default, so admins should decide which kinds of meetings Claude may join. 📎 Read the original
🔟 Google releases new model "Gemini 4 Hydrogen"
🏷️ AI models, Pricing, Google What happened? On November 12 (US time), Google released a new model, "Gemini 4 Hydrogen," and positions it as a major upgrade for developers moving from Gemini 3.8 Flash. Google says that with the default setting (thinking depth "high") it finishes work about 2.5 times faster than 3.8 Flash. The API price is $1.50 per million input tokens and $7.50 per million output tokens. On the 10th, Google also opened its top model, "Gemini 4 Carbon," which had been limited to cyber defense organizations, to all paying API users. Carbon costs an introductory $2 for input and $10 for output until November 30, then $4 and $20 from December. Our view Google gave Carbon to defense organizations first and opened it to general developers about two weeks later. Companies that started using Carbon through the API should recalculate their bill at the December prices, and also compare the cost and quality of doing the same work on Hydrogen. 📎 Read the original
📚 Editor's Note
According to OpenAI, ChatGPT Wallet does not pass the user's card details to the agent. According to CNBC, what the agent uses is a set of card number, security code and expiry date issued by the card network or similar, which works only for that transaction.
This week, as with Notion's contractor and Grok Build, there were several cases where data or sharing links around AI stayed usable from outside. When you choose an AI service, check whether the keys and card details you hand to AI can be made usable for that one purpose only.
See you next week, with useful information and something to think about. The dera news team