New models
GPT-6 Cyber: OpenAI's cybersecurity model is available and crosses designated boundaries less often in tests. This is the model behind the Linux vulnerability research shown yesterday, and OpenAI says it found 81 new vulnerabilities in Firefox. On a deliberately trap-built test, the rate of touching systems outside the task fell below 1% from the previous model's 5%. For now only companies in the Daybreak Red program get it, with access expanding gradually. Source: OpenAI (X), The Verge, TechCrunch.
Gemini Omni 1.2 Flash: 15-second videos 30% cheaper with Google's new video model. It makes 15-second clips instead of the previous 10, follows reference images better, and the video can be edited in conversation across multiple steps. It is available in the Gemini app, AI Studio and the Gemini API, as well as on OpenRouter and Runway. Source: Google, Google AI Studio (X).
ERNIE 5.5 Flash Base: Baidu releases an open foundation model built for further training. The 280-billion-parameter MoE model (14 billion active per token) freezes weights at the end of pretraining, so developers decide which direction to train it further. It handles a one-million-token context during fine-tuning and inference. Source: Baidu ERNIE, The Decoder.
Ling-3.2-flash: Ant Group unveils an agent model that also understands images, documents and 3D. The 560-billion-parameter model (25 billion active) beats Claude Opus 5 on several document- and image-understanding tests by the company's own measures. Demos show it coordinating multiple agents to produce analysis, a presentation and video, and it also drives Blender. Source: Ant Ling (X).
Cosmos 3: Nvidia releases four open world models for simulating self-driving cars and robots. The 16-billion-parameter models generate, among other things, multi-camera driving footage in varied weather, synchronized camera and LiDAR data, and video predicted from robot motions, under a license that also allows commercial use. Source: Hugging Face: Nvidia.
Ampersand: a coding model from Humans& trained to work alongside people. The company says that when run as multiple agents it matches leading coding models while being faster and cheaper, even though it received no separate coding training. Source: Humans& (X).
Tools & features
Claude Personal Finance: bank accounts can be linked to Claude, which can only read them. The beta for US Pro, Max and Ultra subscribers connects to bank and credit-card accounts through Plaid and answers everyday questions about spending. It cannot move money, and Anthropic says it does not recommend credit products. ChatGPT got a similar feature in May. Source: Claude (X), TechCrunch.
Perplexity Teammates: continuously running workplace agents that get goals, not instructions. They run in an isolated environment, take over existing projects, skills and memory, and also carry out Perplexity Computer tasks. For now they are available on the web to the most expensive Max subscribers. Aravind Srinivas says about ten thousand such agents already work on the company's internal workspace, twenty times the number of employees. Source: Aravind Srinivas (X), The Verge.
Mistral Studio: supervised agents and apps, company connectors and evaluation on Mistral's developer platform. The update adds workflows, observability and evaluation tools, and where the work runs remains the company's choice. Source: Mistral AI (X).
Cursor: the editor builds and deploys a working, shareable-by-link app from a description. Cursor also sets up a database, runs security checks, then publishes the finished internal tool so it does not stay stuck on the developer's machine. Source: Cursor (X).
Muse Voice: US users can talk with Meta's AI agent by phone and video call. Voice conversations come with a customizable avatar that reacts to the mood of the chat with facial expressions. The Muse app has also launched in Mexico. Source: Meta, TechCrunch.
Private Safety Processing: the OpenAI API runs safety checks without retaining the requests. Logs stay in the customer's own environment, where an automated safety agent reviews them, and OpenAI staff cannot access the content. Source: OpenAI Devs (X).
ElevenMusic: AI music made with ElevenLabs can be uploaded straight to SoundCloud. The company says more than 25 million songs have been made with its models so far. SoundCloud Artist Pro subscribers get the ElevenMusic Pro plan free for three months. Source: TechCrunch.
Developer odds and ends: Hugging Face's machine-learning intern becomes a Slack bot, Antigravity opens Jupyter notebooks. The open-source ML Intern evaluates models, builds datasets and fine-tunes on your own Slack workspace. Antigravity meanwhile renders .ipynb files with their outputs and also got a new interface. Source: Hugging Face (X), Antigravity (X).
Policy & safety
METR: at Anthropic, model progress is 1.9× faster than before October 2025, and AI itself largely drives that. The independent evaluator has studied Anthropic's internal data since August to see whether AI speeds up its own development. They say it probably does, and they see no clear brake over the next 12 months. METR also objected that Anthropic redacted most of what they considered the most important evidence from the public version. Based on the report's data, the company's internal model finished in October (Model 4) sits well above Opus 5.5 on their own capability index. Source: METR (X), TestingCatalog (X).
OpenAI: will not sign NDAs with local decision-makers on its data center projects. After election losses, the company made new commitments to communities hosting data centers: it would use less power and promise residents a say and a share of the upside. Confidentiality would cover only genuine security and trade secrets. Source: OpenAI Newsroom (X), The Verge, Bloomberg.
Claude Code: a developer says it sent their email address to pages visited during web search. According to a GitHub bug report, the address ended up in the User-Agent header of requests, and the user has since received dozens of unsolicited emails a day. The report drew hundreds of confirming reactions; Anthropic has not yet issued an official response. Source: GitHub: claude-code #107757.
US Congress: Scott Wiener, author of California's AI safety law, is in, and a comprehensive federal AI law is forecast for spring. Wiener wants federal AI oversight and would expand unemployment benefits for people who lose jobs to AI. Brad Carson, head of an AI safety advocacy group, says the strengthened Democratic majority could pass a law covering the strongest models by spring. Source: The Information, The Information (X).
Meta: indefinitely suspends new business terms that would have allowed scraping corporate customers' websites. After publisher protests and court orders, the changes planned for November 20 will not take effect, and the old terms remain in force. Source: Nieman Lab.
Chip exports: about 600,000 Nvidia H200 chips sit in Hong Kong because Beijing will not approve large-volume imports. Meanwhile three Republican House committee chairs want to question Jensen Huang in closed session about his ties to a businesswoman they consider a Chinese agent. Source: Reuters, CNBC.
ChatGPT lawsuit: a man says his wife used the chatbot to plan his murder. The complaint says that instead of warning or ending the conversation, ChatGPT calculated lethal drug doses and advised how to administer them unnoticed. Source: Ars Technica.
Europe: more than 100 groups in 15 countries protest against data centers on Friday and Saturday. The "Stop Dirty Data Centres" action demands that large projects be halted over environmental and social costs. Source: The Guardian.
Business & society
Markets: the Nasdaq 100 fell 6% with a record $2.3 trillion wipeout, and the chip index plunged 12%. Investors fear data center construction will slow after the Democratic win, so a day after big tech rose, the entire AI trade reversed. Memory maker SanDisk fell 22%, and Asian markets also plunged on Friday. Source: Bloomberg, The Guardian.
OpenAI: Greg Brockman says the company could be IPO-ready in 2027, and it struck a strategic partnership with JPMorgan Chase. Brockman says the first half of the year was "pretty hard," but annualized revenue grew 70% in the third quarter, and enterprise customers already provide most of the revenue. With JPMorgan they are collaborating on cybersecurity among other areas. Source: TechCrunch, Greg Brockman (X).
Data centers after the election: Meta says new ones will be harder to build, Google is spending $17 billion in Wyoming, Microsoft delayed $5 billion of investment. Meta's president says expanding existing projects will be easier than opening new sites. Google pledged to pay for the Wyoming data center's full power use and to set up a local advisory council. Microsoft is supervising leased capacity more tightly because of delayed partner projects. Source: TechCrunch: Meta, Bloomberg: Google, Bloomberg: Microsoft.
Anthropic: pays for news content for the first time, in a deal with the AP. According to Adweek, the agreement let Claude show live US election results with source attribution. Anthropic had previously refused such licensing deals by citing fair use. Source: Adweek, The Decoder.
Dropbox: buys AI storage company Tigris Data for $350 million. It is the company's largest acquisition so far; the new DBX Agent Cloud service first offers S3-compatible object storage for AI agents and apps. Source: TechCrunch.
Airbnb: its own AI travel agent launches in the first half of 2027. Brian Chesky says the travel-planning experience should not be swapped for a chat window, so they are designing a richer interface. It also emerged that Airbnb received an option to invest in Chesky's own-funded AI lab. Source: TechCrunch, The Information.
Agility Robotics: the first US company focused purely on humanoid robots goes public. Listed on Nasdaq as AGLT, the company itself acknowledges in its filing that mass use of humanoid robots is not yet proven. Source: Bloomberg, Ars Technica.
Cat Wu, former product lead of Claude Code, joins OpenAI. The executive who left Anthropic last week now leads product management for OpenAI's core products. Source: Business Insider, Greg Brockman (X).
Prix Goncourt: this year's winner of the French literary prize admits using ChatGPT on the novel. Philippe Jaenada says he slipped one ChatGPT sentence into the text for fun, and in one passage he mocks the chatbot's helpless answers — a day after the jury excluded another book on suspicion of AI use. Source: The Guardian.
Rumors
Palo Alto Networks: may buy Clutch Security, which oversees AI agents' password use, for $500 million. According to The Information, talks are underway; there is no confirmation. Reliability: medium-high (unnamed sources, but usually accurate business reporting). Source: The Information.
Nvidia: would turn old telephone-exchange buildings into small data centers to dodge protests. According to The Information, Nvidia proposed to near-bankrupt Brightspeed that more than 130 buildings be converted. Reliability: medium (single source; the proposal has not yet become a deal). Source: The Information.
What this means for you: Thursday brought both agents that work on their own and warning signs: Perplexity, Mistral and Cursor all released continuously running or self-deploying agents, Claude can already connect to bank accounts, a bug report says Claude Code leaked a user's email address, and a comprehensive AI law may be coming in Washington. If you would let agents into your company's systems, on our AI agents page, you can see which tasks they can take over under controlled conditions, and with our AI governance and compliance support, you can set in advance what data they may touch and what must be logged.