Tools & features
Nous Research: Hermes Agent gets a security plugin that hunts for vulnerabilities. The new official Security plugin first maps what an attacker can reach and which inputs the code trusts, then splits the codebase across parallel subagents, checks every suspicious point against the code and tests whether it is truly exploitable. With test confirmation on, it writes a small reproduction for every finding and runs it in a temporary copy of the repo; the report records both the evidence and the ruled-out findings, and can be exported in SARIF into existing security tools. It can be started from chat, the command line or the Hermes Desktop Security page. Source: Nous Research (X).
OpenAI: none of the Pro plans has a five-hour limit, the Codex lead clarified. According to Thibault Sottiaux, Pro 100 gives ten times the Plus usage, Pro 200 twenty times and Pro 500 fifty times, and the latter also unlocks "ultrafast" mode. In a separate post he also asked users whether a Pro 50 plan would make sense. Source: Thibault Sottiaux (X).
Lovable: a PayPal connector for apps. An app built with Lovable can now create a payment with an existing PayPal account, collect the amount after the buyer approves, issue refunds and read order details — for a webshop checkout, a fundraising page, ticket sales or a booking deposit. It can be turned on from the Connectors menu or by asking Lovable in the project. Source: Lovable changelog.
Policy & safety
Nvidia: Jensen Huang admitted an affair, and the board opens an independent probe into the spy case. In a Sunday letter to staff, the CEO wrote that he had been unfaithful to his wife, took full responsibility and acknowledged that he had not lived up to his duty to the company; he did not name the woman. According to press reports, US intelligence believes Alice Huang, head of Singapore-based Megaspeed, was recruited by China's Ministry of State Security. Independent board members are investigating with outside counsel whether Huang was the target of a foreign intelligence operation; they say they have no reason to believe national security or company secrets were compromised, and they trust Huang's leadership. According to The New York Times, a classified memo prepared under Marco Rubio warned the Trump administration about the relationship in the spring; Trump on Sunday called Huang an honest man who should stay at the top of the company. Source: Nvidia, TechCrunch, Bloomberg, CNBC.
Staff at CAISSI, the US AI safety center, have been asked to report informally to the White House. Citing two former officials, AP writes that the Commerce Department center that evaluates advanced model safety has no permanent director and has stopped publishing its research findings. The director appointed in April was dismissed within days, and his successor resigned after a few months; the annual budget is $10 million, roughly one eighth of its UK counterpart. Critics say safety decisions could become politicized; others say proximity to the White House could bring more attention and funding. The White House did not answer the specific questions. Source: AP.
Tencent: models' biological capabilities will soon reach a level where they can cause real harm. At a Hong Kong science symposium, Yao Shunyu, Tencent's lead AI researcher, said that after cyber capabilities, biological capabilities will also reach this threshold and may be harder to manage. He said Anthropic has a "strong case" for closed models, but the best mix may still be open and closed together, because open weights also help defenders; Chinese companies releasing open models must take safety very seriously. Tencent left offensive cybersecurity material out of the training data for its Hy4 model unveiled this week. Source: South China Morning Post.
The IFPI warns Australia against weakening copyright for AI companies. IFPI CEO Victoria Oakley told Guardian Australia it would be a "very strange decision" if the Albanese government gave in to big AI companies' requests and opened the door to training on online content that other countries have already rejected. Australian creators have also raised constitutional concerns and threatened a political backlash. Source: The Guardian.
Business & society
Anthropic: bought 1.7 gigawatts of gas generators for its data centers for $3.4 billion. According to the Financial Times, in late September Anthropic took over Texas-based OnePWR Solutions' prior Caterpillar orders, together with the engineering and construction capacity needed to install them; it did not buy OnePWR itself. The paper says this is the largest known power-generation equipment purchase by an AI lab. According to The Information, Anthropic is also looking for 5–10 megawatts of capacity in old telecom buildings because of a server shortage. Source: CNBC, The Information (X).
Apple: John Ternus is pushing for a faster, more agent-like Siri. According to Mark Gurman (Bloomberg), the CEO is urging Craig Federighi to bring Siri AI up to the level of Dots- and Muse-like agents. With iOS 27.3, Siri will more finely control Apple's own apps — for example crop a photo in Photos or rearrange Safari tabs — third-party app integration may arrive in the spring, and Gurman says that as soon as next year Siri may independently chase longer, multi-step goals. Source: Bloomberg, The Verge.
DeepSeek: its open-source Harness agent app is at 1.54 million daily users. Missed in our previous roundups: according to the team, daily active users measured through the official API hit 1.54 million on November 4 (a seven-day average of 271,000 at the end of September), they use about 92,000 trillion tokens a day, and roughly 35,000 external plugins have been built. An official plugin marketplace is planned; Bloomberg reported the figures on Sunday. Source: Bloomberg, TechNode.
AI companies have sold or priced more than $300 billion of stock this year, not counting Anthropic's IPO. According to The Information, roughly another $100 billion of issuance is in preparation, raising the question of how much more the market can absorb. Source: The Information.
Despite fears of AI, professional and business services in the US added 70,000 jobs in October. According to a Wall Street Journal analysis, that is the largest monthly gain since July 2022, and consulting, legal and accounting also grew; employment in document preparation, however, was down 17 percent year on year through September. Source: The Wall Street Journal.
Co-Design Bench: on real user requests, even the best coding agent solves only 26.4 percent of tasks without errors. Researchers from Alibaba and several Asian universities collected modification requests from real coding sessions on which agents had previously stalled. Of the ten models tested, leading GPT-6 Astra meets about 70 percent of the requirements on average, followed by Claude Opus 5 and Qwen3.8-Max, while on older tests these same models score above 99 percent. The more kinds of change a task asks for (layout, design, interaction, content), the weaker they are. Source: The Decoder.
“AI clog”: British public bodies are wrestling with a flood of chatbot-written complaints. According to The Guardian, MPs' offices, local authorities, lawyers and schools are receiving more and more AI-written complaints, requests and legal filings; one constituency office aide says half the letters are AI-generated in some form, and processing them takes far more time and energy. Source: The Guardian.
Amazon: a new team is working on a home robot. Based on job postings, CNBC writes that Amazon is developing a "world-new" robot product that would operate independently at home alongside family members — after the $1,450 Astro reached only a few thousand homes. Source: CNBC.
SemiAnalysis: an on-site report from China's “Stargate” in Inner Mongolia. Missed in our previous roundup: the analyst firm visited a wind-powered 800-megawatt campus where 120 megawatts were built in a year, a 5-megawatt solid-state-transformer 800-volt DC pilot system is running, and cooling works without central chillers. The region is China's fastest-growing AI data center area, with low construction costs and no local resistance. Source: SemiAnalysis.
Rumors
GPT-6.2 Sol: some ChatGPT users may already be getting the new model, and OpenAI Developers posted one word: “Tomorrow.” According to several user tests and kimmonismus, part of GPT-6.1 Sol requests is being served by a newer model that the community calls GPT-6.2 Sol; OpenAI's developer account posted a single word (“Tomorrow”) on Sunday evening. OpenAI has not confirmed the model's name or contents. Reliability: medium (consistent user tests and an official teaser, but no specifics). Source: OpenAI Developers (X), kimmonismus (X).
Google DeepMind: a “phase transition” in scientific discovery, with releases coming in the next few weeks. Pushmeet Kohli, Google DeepMind's head of scientific research, says their projects are seeing a real phase transition in the pace of scientific discovery, and they will show a taste of it in the coming weeks' releases. He shared no details. Reliability: medium (an official leadership teaser, no specifics). Source: Pushmeet Kohli (X).
Meta: Alexandr Wang promises a Muse launch “this week.” The Meta AI lead pointed to a weekly announcement with a video of dancing Muse avatars; exactly what launches was not clear. Reliability: medium (an official leadership teaser, no details). Source: Alexandr Wang (X).
Google may be working on a next-generation Gemini agent for Gemini Enterprise. According to testingcatalog, Google is already using an interface named “Gemini for Google – The next generation Gemini agent for enterprise” internally. There is no official announcement. Reliability: medium (a single leaker who has been accurate several times before). Source: testingcatalog (X).
What this means for you: on Sunday, security and infrastructure dominated: Nvidia's CEO spy case, the leaderless US AI safety center and the labs racing for power all show that choosing an AI supplier is also a risk-management question. If you would entrust code or a business process to an agent, on our AI agents page, you can see which tasks are worth handing over, and with our AI governance consulting, you can set supplier and oversight rules in advance.