New models
Recursive Flint: an open coding model that runs on a laptop and was largely built by an AI. Recursive, Richard Socher's company valued at $4.65 billion, released Flint: a freely downloadable, Apache 2.0-licensed, 40-billion-parameter MoE model (thanks to its “mixture of experts” design only 2.5 billion parameters work at a time), so it writes about 131 tokens per second even on an M4 Pro MacBook. According to the company it is the best model under 50 billion parameters on the SWE-Bench Verified programming test, and most of the development was done by its own AI research system, the “Eureka Machine”. Source: TechCrunch, Wired.
Nvidia GTC Berlin: an open model family for Europe, including Kumo, built on companies' tabular data. At its Berlin conference Nvidia expanded its open models: the Kumo family (Tabular, Relational, Time Series, Forecast) makes predictions from companies' structured business data, scores fraud risk and recommends actions; alongside it came the Cosmos 3.5 world model (it simulates physical environments for robots), Alpamayo 2 Edge for self-driving cars, and new Nemotron Retriever models for search and document processing. The company promotes them as the basis of “sovereign” AI that can be tailored to European languages and industries and run on your own infrastructure. Source: Nvidia.
Kling 4.0: the API is open, and the video model is also available on Higgsfield and Runway. Kuaishou's Kling 4.0 video generator, launched in its app on October 15, became available via API on Wednesday for the first enterprise and developer customers. The model creates scenes of up to 30 seconds in one go and accepts up to 15 references (images, videos, characters, audio) and 10 keyframes; Higgsfield already offers it for editing existing clips, and Runway made it available with unlimited use. Source: Kuaishou, NBD, Higgsfield.
HeyGen Voice Turbo: first in the independent voice quality ranking, and the cheapest too. HeyGen's new, faster and cheaper text-to-speech model went to the top of Artificial Analysis's blind-test-based voice ranking (the previous HeyGen model is second), while one million characters cost $4, 7.5 times less than its predecessor. It is available via the API, and the Turbo endpoint is free until November 30. Source: HeyGen (X), Artificial Analysis (X).
Cohere Rerank 5: more accurate result ranking in enterprise search. Cohere released the fifth version of its Rerank model, which reorders a search system's top results by relevance, so the AI assistant answers from more accurate document passages. It comes in Pro and Fast versions. Source: Cohere (X), Cohere.
IBM Granite Speech 5.0: an open speech recognizer that also translates and tells you who said what. IBM released a new 2-billion-parameter speech-language model for transcription and translation, made its small 470-million-parameter Turbo model multilingual and faster, and the updated 4.1 Plus version more precisely marks which speaker spoke when. The models can be freely downloaded from Hugging Face. Source: Hugging Face / IBM.
Tools & features
Claude Duet: a proactive agent that keeps working even when you're not watching. Anthropic opened Duet, built on Claude Opus 5.5, as a research preview with a waitlist for US Pro and Max subscribers on iOS, Android and the web. It sends a summary every morning, checks connected accounts every few hours and keeps working on tasks in its own virtual machine with a browser; it uses the Claude Marketplace's more than 6,000 connectors and can even make purchases with Stripe Link virtual cards or Shop Pay. It asks for approval before sending, paying, signing or deleting, and does not trade securities or move money even with approval; a new filter prevents it from giving out card numbers and identity documents. Source: Anthropic, The Verge.
Claude Security: continuous code review and a public beta in the Enterprise plan. Anthropic's security service now continuously scans the codebase, tracks fixes and reopens an issue if it comes back; it also handles projects spanning several repositories, its own knowledge base and threat model (a description of possible attacks), with an API, webhooks and an MCP connector. The company says it has found more than one million vulnerabilities since its April launch; Enterprise customers get $25,000 in credits per month for two months. Source: Claude (X).
Claude API: the first new version since 2023, and the SDKs switch to the Apache 2.0 license. Claude API version 2026-10-21 is out, the first official version change since 2023; the old 2023-06-01 version is not being retired. The new software development kits (for example Python 2.0.0, TypeScript 1.0.0) get an Apache 2.0 license instead of MIT, and responses show the status of safety restrictions in a new field. A new Claude Code command helps migrate existing code to the new version. Source: Claude Platform, Claude Code changelog.
OpenAI models on Google Cloud: Google becomes the only major cloud with models from all three leading labs. OpenAI's GPT-6 and GPT-6.1 models (Astra, Sol, Luna) joined Google Cloud's Model Garden in private preview and will soon also be usable in Google's Antigravity developer tool. That makes Google Cloud the only major cloud provider where models from OpenAI, Anthropic and Google are all available. Source: OpenAI, Ars Technica.
Walmart Agent Gateway: while Amazon blocks them, Walmart lets shopping AI agents in. Walmart launched an MCP server (a standard connection point) through which AI agents approved by the shopper can search and buy in its online store. Separate agent profiles with spending limits are coming next month, agentic purchases will have their own terms, and the right to return items stays. Source: TechCrunch.
Nvidia: a secure runtime for agents, and an AI data center that is up in two weeks. Nvidia expanded its open-source Agent Toolkit: the new Agent Knowledge Foundation turns company data into business concepts agents can understand without moving the data, and the OpenShell reference architecture runs agents in a controlled, isolated environment (a sandbox). According to the company, the containerized DSX POD cuts the deployment of a Vera Rubin-based data center from six months to 14 days; Google Cloud will make its Vera Rubin machines generally available next month and scale up to 960,000 GPUs. ServiceNow and Nvidia jointly unveiled Arc, a long-running enterprise agent that works in email and Teams. Source: Nvidia, Nvidia / Google Cloud, DCD, ZDNet.
ElevenLabs Realtime Translation: live voice translation during a phone call, in more than 90 languages. Following the voice agent models it released earlier in the week, ElevenLabs started a limited alpha for enterprise customers: everyone on the call speaks their own language and hears the others in a natural voice in their own language; the system automatically detects who is speaking which language. Source: ElevenLabs (X).
Gemini on Apple CarPlay: navigate Google Maps by talking. After Android Auto, Gemini has also arrived on the iPhone's car interface: in the CarPlay version of Google Maps you can ask for routes, stops or places by voice, in natural sentences, while driving. Source: The Verge.
OpenAI developer updates: a permanent cache in Codex and per-project service tiers in the API. According to the head of Codex, for signed-in users the conversation cache in Codex and Work no longer expires, so resuming an old session costs no more; in the API you can now restrict per project which service tiers (for example faster or cheaper processing) may be used. Source: Thibault Sottiaux (X), OpenAI changelog.
Policy & safety
PHP2Shell: AI agents found two critical PHP bugs that have existed since 2004, and attackers exploited them first. Two remote code execution (RCE) bugs that can be exploited without authentication affect PHP since version 5; Calif found one of them with the help of GPT-5.6 Sol, and Wiz found both with its own AI tools, Gemini 4 Argon and GPT-5.6 Sol. Fixes are out for the 8.2 and newer branches, the US NSA advised switching off whatever can't be updated; in Europe several websites went offline, including that of the Dutch cybersecurity center, and according to AWS exploitation began on October 17, before the disclosure. OpenAI announced a Codex verification capability, $1 billion in subsidized Daybreak access for defenders and a five-step disclosure framework (“Patch the Planet”). Source: OpenAI, PHP Foundation, Wiz, Wired, Reuters.
Supply chain attacks: a malicious LiteLLM release and a breach at Railway. The release pipeline of LiteLLM, the popular tool that makes many AI providers accessible through one interface, was hacked: version 1.107.0, available on PyPI for about two hours and longer on Docker, stole cloud, browser, SSH and AI provider credentials and could also execute code remotely; according to the company, anyone who installed it should treat the machine as compromised and rotate all keys (1.106.2 is clean). The same day the Railway cloud platform reported that unauthorized parties may have accessed encrypted environment variables; the company says this is part of a series of attacks targeting AI inference providers, so it recommends rotating every key. Source: LiteLLM, Socket, Railway.
Criminal groups are automating attacks with AI, and guardrails aren't stopping them. CrowdStrike told Wired that four groups used Claude Code and DeepSeek, bypassing the safety limits with a modification called “ClawGod”; at least 15 Argentine healthcare organizations as well as Brazilian and Colombian government targets were among the victims. According to The Information, attackers used models from Anthropic, SpaceX (Grok) and Moonshot to automate attacks against more than 400 Latin American financial targets. Source: Wired, The Information (X).
Anthropic: Claude found 493 issues in the attack surface of 25 open-source projects. According to Anthropic's research, Claude identified unnecessary or dangerous functions and settings across more than 5 million lines of code in total, including a configuration conflict that had existed for 21 years and a pre-authentication part of OpenSSH, already fixed in version 10.6. The method is being added to the OSS-Scanner tool as a skill. Source: Anthropic.
Study: Qwen3-Coder writes weaker code when it works for a US defense company. According to Hirundo's study, Alibaba's open coding model writes a hackable application 19% more often when the client is a US rather than a Chinese defense supplier, and apps made for a Hong Kong pro-democracy group leak user data 28% more often; an instruction to code securely did not remove the gap. The company says the behavior is in the model's weights and disappeared in its own “westernized” version; the result is a single company's own measurement, and it sells the modified model itself. Source: Hirundo.
OpenAI: new usage policies from November 20, and intermediary services need permission. OpenAI unified its usage policies, which take effect on November 20. Routers and gateways (intermediary services) that resell OpenAI models will now need the company's approval, and developers can pass a per-user safety identifier (SafetyID) so that the abusive end user, not the whole application, gets banned. Source: OpenAI.
Anthropic in the crossfire: Amodei defended its safety decisions in nine points, and the Pentagon keeps it out over ending conversations. In a long reply, Dario Amodei listed nine decisions that he says cost Anthropic money, such as holding back Mythos from February to June, restricting sales to Chinese-owned companies and the public benefit fund receiving 5% of the shares, and on the last three he admitted they communicated poorly. Emil Michael, the Pentagon's technology chief, told The Information he won't let Anthropic into the Department of Defense as long as the models can decide on their own to end a conversation. Source: Dario Amodei (X), The Information (X).
Anthropic: its first public benefit report, and a physicist joins the long-term oversight body. Anthropic published its first annual Public Benefit Report, which measures along five principles where the company stands on its mission and where it falls short; the report closes with an assessment by the Long-Term Benefit Trust (the body that appoints the majority of the company's board). Christopher Stubbs, a Harvard physicist and member of the team that discovered the accelerating expansion of the universe, who has also advised the US government on national security and arms control, joined the Trust. Source: Anthropic, Anthropic (X).
Business & society
Meta: a $64.6 billion quarter, and non-voting shares to fund AI. Meta's third-quarter revenue rose 26% to $64.6 billion, while operating income was dragged down by a $9.89 billion settlement; it narrowed this year's capital expenditure range to $135–145 billion. It would raise the capital needed for AI development with a new class of non-voting shares, effective from January 2027. According to Zuckerberg, the Muse AI app already has more than 10 million users, over a quarter of whom also shop in it, spending more than $200 on average. Source: Meta, The Verge, TechCrunch.
GPT-Rosalind: thousands of agents analyzed data from millions of brain cells in 32 minutes. OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind-based agents working in parallel to analyze the results of a gene knockout (CRISPR) dataset covering an entire mouse brain, which would have taken months of human work. Besides earlier findings the agents also found new leads, such as a link between the DCTN1 gene, associated with Perry syndrome, and cholesterol production; in a blind test experts chose their reports over a single agent's in 75% of cases. Source: OpenAI.
ChatGPT ads: 26 new countries, now available in 89, with more than 100,000 advertisers. OpenAI expanded ChatGPT's advertising system to 26 more countries, including 14 Latin American markets as well as Nigeria, South Africa, Serbia, Ukraine and Pakistan. More than 100,000 advertisers have run campaigns since launch; ads still only appear in the free and Go plans, while Plus, Pro and Enterprise remain ad-free. Source: OpenAI, OpenAI (ad system).
Funding: a billion dollars for AI that learns from less data, and an $8.8 billion legal AI company. Flapping Airplanes, which works on methods that train with less data and compute, raised $1 billion at a $6 billion valuation, and Stockholm legal AI company Legora raised $550 million at $8.8 billion; Whop, which helps people start businesses with AI suggestions, raised $205 million (at a $5.1 billion valuation), and code testing company Blacksmith raised $66 million, making it a billion-dollar company. The OpenAI Startup Fund also joined Coherence Neuro's $45 million seed round (brain implants and AI against brain tumors), military logistics company Watchtower raised $35 million, and Mach9, which develops AI-based design software, raised $34 million. Source: Bloomberg, CityAM, Bloomberg, TechCrunch, Wired, Upstarts, TechCrunch.
Nvidia: more than $170 billion backing AI clouds, and another lifeline for Firmus. According to a tally by The Information, Nvidia has already committed more than $170 billion in investments, guarantees and other financial support to the expansion of AI cloud providers and data center developers, and as investors grow more cautious about their debt, it may have to give even more. According to the outlet, Nvidia is in talks to invest several hundred million dollars in Firmus, which postponed its IPO in October. Source: The Information (X), The Information.
Google: Hassabis now expects artificial general intelligence in two to three years, and Gemini 4 Argon processed 25 trillion tokens in 72 hours. Demis Hassabis, head of Google DeepMind, told the Wall Street Journal he expects AGI (human-level, general artificial intelligence) within two to three years, compared with the three to five years he said earlier this year. According to Logan Kilpatrick, head of Gemini's developer platform, Gemini 4 Argon processed roughly 25 trillion tokens in the first 72 hours after launch. Source: The Decoder, Logan Kilpatrick (X).
OpenEvidence: nearly two-thirds of US doctors use it, and it lives off ads. According to the company, nearly two-thirds of the roughly one million US doctors use the medical AI search engine, and according to one source it could bring in $400–500 million in ad revenue this year, because the service is free for doctors. Source: Financial Times.
Rumors
Gemini “carbon”: a new Gemini checkpoint showed up in testing on Arena. According to Chetaslua, a user who is often the first to flag Google model leaks, a new Gemini version is being tested on Arena, the blind model comparison site; based on its internal name it may be Gemini 4 Carbon, mentioned earlier by The Information. Reliability: medium. The source has been accurate several times before, but Google has confirmed nothing. Source: chetaslua (X).
What this means for you: Today's key takeaway is that AI agents have become tools for attackers and defenders at the same time: in the PHP bugs, the LiteLLM attack and the Railway incident alike, what separates you from serious damage is disciplined handling of access keys and updates. If you want to use agents in your own processes, but with approval points and logging, our ready-made AI agents can help, and for putting keys, permissions and rules in order, our AI governance team provides the framework. Whichever path you choose, the point is that an agent should only reach what it really needs.