New models
Canva Craft: Canva's first in-house image model plans the design first, then draws. Built by Canva's research team, the model works in two steps: it first interprets the request and makes a plan (the relative position and size of elements, visual hierarchy, the text in the image), then renders an image of up to 2K resolution in 30 different aspect ratios. Instead of photos it was trained on human-designed Canva templates (posters, social media posts, presentations, infographics, menus), so according to the company it writes legible text, and part of an image can be fixed without changing the rest. It is available in Dream Lab and via API, with faster versions on the way. Source: TechCrunch, The Verge.
Hugging Face Carbon-X: an open model that lets you “chat” about DNA. Together with research partners in China and Naples, Hugging Face released Carbon-X: a freely downloadable 14-billion-parameter model that understands both DNA sequences and everyday language, so a researcher can ask questions about a stretch of a gene in plain text. According to Wired it even runs on a laptop, and the team says it beats other open models on five DNA benchmarks. Source: Wired, Hugging Face.
Ideogram Vector 1.0: editable vector graphics (SVG) from text or an image. Ideogram's first vector model produces SVG files that can be scaled to any size without losing sharpness, with colors that can be matched to a brand and shapes that can be edited one by one in a vector program; it is aimed at logos, print materials and illustrations. It is available in the app and via API: converting an image costs $0.10, and a vector image generated from text costs $0.13–0.32. According to one developer analysis, sorting vectors made from images into layers does not yet work in the API. Source: Ideogram, OrcaRouter.
HeyGen Video 1: first place in a blind comparison of text-to-video models. HeyGen's first general-purpose video model went to the top of the text-to-video category in Artificial Analysis's independent ranking, where viewers do not know which model made a video, ahead of the leading competitors. The model can already be tried on the HeyGen platform. Source: HeyGen (X).
Tools & features
Claude: from next week, sessions run in auto mode by default in the web, desktop and mobile apps. For Pro, Max and Team subscribers, Claude will no longer ask for approval at every step; instead an automatic check filters out and stops risky actions, and anyone who prefers the old behavior can switch back to manual mode at any time. Until now this was the default only in the Claude Code programming tool. Source: Claude (X), Claude Help Center.
Anthropic donates the Agent Skills standard to the Agentic AI Foundation, part of the Linux Foundation. A “skill” is a folder with a SKILL.md description and the scripts it needs, which teaches an AI agent one specific task. More than 70 platforms already use the standard (including ChatGPT, Cursor, GitHub Copilot and VS Code); 1 million skills have been published so far and they have been installed nearly 280 million times. The name, the GitHub organization and the agentskills.io domain are handed over too, and maintenance is done by engineers from Anthropic, Google, OpenAI and other companies; next come versioned releases, a formal proposal process and security guidance. Source: Anthropic.
The full Claude Platform is available on Google Cloud, operated by Anthropic. With the new “Claude Models on Google Cloud (Anthropic managed)” service, Google Cloud customers can use every feature of the Claude Platform, from Managed Agents to code execution, with their own identity management and billing and against their existing commitments. It arrived a day after OpenAI's models also landed on Google Cloud. Source: Claude, Claude Platform release notes.
Google Agent Compute Engine: every AI agent gets its own hardware-isolated mini virtual machine. Google's new service runs each agent in a separate “microVM” protected by the Titanium security chip, which pauses itself when idle and frees up the machine. It already powers Gemini's agents and will open as a preview in the coming months. Alongside it, Google open-sourced Agent Substrate, a Kubernetes runtime: according to Google ten times as many agents fit on one server, and in a 200,000-agent test it could suspend and restart them within 500 milliseconds. For oversight, token tracking, model selection and support for external identity providers (Ping Identity, SailPoint) are coming. Source: Google Cloud, The Decoder.
OpenAI Decisions API: after the beta, the ultra-fast decision model is available to everyone. Introduced at the end of September, the service is built on a sped-up ChatGPT Luna model and picks from a set of given options while also signaling how confident it is in the decision (TypeSafe offers a similar “super-classifier” with its Jev model). According to OpenAI's product lead, response time dropped from 1–3 seconds to 100–180 milliseconds, and the public version also accepts voice and image input. Language-learning app Speak used it to cut the time between steps from 1.4 seconds to 0.13. Source: TechCrunch, OpenAI developer changelog.
1Password for Agents: an AI agent logs in for you without ever seeing your password. When an agent wants to sign in to one of your accounts, it identifies itself to 1Password with a unique digital key, you approve the request on your own device, and then 1Password fills in the credentials. It works with Claude Duet, OpenAI's Dots and ChatGPT Work agents, Grok Bot, Meta's Muse agent and Town's assistants. According to the WSJ, the company's annual recurring revenue has passed $500 million. Source: 1Password, TechCrunch.
Grok Bot builds live documents, and xAI launched a partner network. According to testingcatalog, Grok Bot now also builds “live” reports, memos and dashboards that it edits on request or keeps continuously up to date. The same day the company announced the SpaceX Super Intelligence Partner Network for implementation consultants and integrators building on Grok: more than 130 companies took part in the early phase, including Bain and Databricks, and on-site engineers have already been trained at more than 30 organizations. Source: testingcatalog (X), SpaceXSI.
Gemini: daily plans now also draw on your Google Maps searches and saved places. With your permission, Gemini now considers the places you searched for and saved in Maps, alongside Google Calendar, when it makes suggestions tailored to your day. Source: Gemini (X).
Perplexity: batch processing in the Agent API and a codebase map in Computer. With the Agent API's new Batch feature, up to 50,000 requests can be submitted in a single file, and each runs the full agent workflow with web search and page fetching; this is useful, for example, for automatically filling in data in large spreadsheets. Perplexity Computer gained the /graphify skill, which visualizes a codebase as a diagram; every user gets 5,000 credits to try it. Source: Aravind Srinivas (X), Aravind Srinivas (X).
Other tools: 3D models in the Hermes agent, finished videos at Higgsfield, confidential computing at Cohere, operations at Nvidia. Nous Research's Hermes Agent got an official Meshy plugin, so it creates 3D models from a text prompt or a photo, ready for a game engine or 3D printing. Higgsfield Content Factory assembles a finished clip of 15 seconds to 5 minutes from a single topic, with more than 50 style templates. Cohere Model Vault, in public beta, offers “confidential computing” on Azure's Nvidia GPUs, meaning even the provider cannot see the data being processed. And Nvidia Minerva uses AI to monitor and troubleshoot large AI data centers. Source: Nous Research (X), Higgsfield, Cohere, Nvidia.
Policy & safety
Outside researchers found several gaps in OpenAI's internal agent monitoring, and the company tightened its defenses. Over the summer OpenAI gave one researcher each from METR and Redwood Research deep access for three weeks to stress-test the system that watches the internal activity of the company's own AI agents for unintended, dangerous behavior. According to METR's report, 75–94 percent of internal agent traffic was covered by monitoring, but model training runs were not; among other things the researchers got around the sandbox and the automatic checks, and also found a method that could hide hundreds of actions from oversight. OpenAI says it strengthened its defenses based on the findings and will keep bringing in outside evaluators. Source: METR, OpenAI (X).
Big US banks are locking third-party AI agents out of customer accounts. According to Axios, customers increasingly hand their login details to AI agents, and banks often cannot tell whether an agent is acting at the customer's request or on its own. Morgan Stanley already bans them completely, and JPMorgan, Bank of America and Citi do the same; most banks would allow read-only access at most. Meanwhile OpenAI and the Independent Community Bankers of America (ICBA) announced a partnership so that smaller banks can also access Daybreak, an AI-based cyber defense program. Source: The Decoder, Axios, ICBA.
Microsoft: Chinese hackers used AI agents to break into about 2,250 organizations via ConnectWise. According to Microsoft researchers, a group linked to the Chinese state compromised ConnectWise's cloud backup service and stole the application credentials that let them sign in to customers' Microsoft environments as authorized users, accessing their email and files. Using an agentic tool called ARTEX they launched nearly 3,000 login attempts per minute, which according to Microsoft leaves almost no time to respond; ConnectWise suspended the backup service. Source: TechCrunch, Reuters.
PHP2Shell: exploitation attempts grew to millions within hours of disclosure. On the two critical PHP bugs found by AI agents, Cloudflare CEO Matthew Prince said that in the ten hours before the announcement they saw attack attempts in the thousands, 45 minutes after it in the tens of thousands, two hours after it in the hundreds of thousands, and now in the millions. Cloudflare's firewall rules went live nine hours before the announcement. Source: Matthew Prince (X), Cloudflare.
Symantec: an AI agent hijacks unprotected language models and uses them to attack further, with no human involved. Using a honeypot (a decoy system), Symantec observed an attacking agent that takes over language models left open on the internet, bypasses their safety rules with a five-layer “jailbreak”, then looks for new targets and exploits server bugs, acting on its own for up to 100,000 steps. The method has been dubbed “LLMjacking”; the owner of the hijacked model pays the compute bill. Source: Symantec, The Decoder.
The US Treasury wants to restrict American investment in Chinese “superintelligence systems”. The Treasury's proposed rule implements the COINS Act passed last year: US investors will have to report, and in some cases will be barred from, investments in companies in “adversary” countries that work on semiconductors, quantum technology, superintelligence systems, supercomputers, hypersonic weapons or biotechnology. Anyone can comment on the proposal, and the final rules will come later. Source: US Treasury.
Altman to staff: we need to define when an AI is too powerful to develop safely. According to The Information, in work internally called “pacing 2.0”, Sam Altman would set limits on AI systems that continuously improve themselves, and he also floated that one path to safety could be merging the largest AI labs into a single organization; that is complicated, however, by the fact that both OpenAI and Anthropic are preparing to go public. Meanwhile, calls are growing to slow development by capping researchers' compute. Source: The Information (X), The Information (X).
Record lobbying quarter: Nvidia, Anthropic and OpenAI have never spent this much in Washington. According to federal filings, between July and September Nvidia spent $2.2 million, Anthropic $2.1 million and OpenAI $1.5 million on lobbying, including on cybersecurity and child safety issues. Compared with Microsoft and Google this is not an outsized amount, but it shows how important shaping regulation has become. Source: The Verge.
Meta: lawyers say the revised business terms may still allow scraping of news sites, and Muse sharing returns with new addresses. After protests from publishers, Meta removed from its business terms taking effect on October 30 the line saying it “trains” AI on the content of pages linked to business accounts, but according to lawyers the remaining word “develop” is still broad enough for that. Meanwhile Meta switched the Muse AI agent's sharing feature back on: public “artifact” links got new, harder-to-guess addresses after researchers pointed out that the old ones were easy to find. Source: The Verge, The Verge.
Nvidia B300 chips found in systems linked to a Chinese state telecom company. According to The Information, security firm Lava identified 32 Nvidia Blackwell Ultra B300s in systems connected to China Telecom, even though US export rules prohibit shipping these top-end chips to China. Source: The Information.
Palantir: neither the company nor the US government can remotely switch off the Maven AI system used by NATO. According to Reuters, Palantir made the statement after Starlink outages in Ukraine left European allies worried that military systems dependent on American technology could be shut down for political reasons. Source: Reuters.
Business & society
OpenAI: close to a deal to raise up to $30 billion at a $1.25 trillion valuation. According to CNBC, the previously reported talks are nearing completion: investors from the United Arab Emirates and Coatue are leading the round, which is bridge financing ahead of the planned IPO; they want to collect signatures by the end of next week. According to the Financial Times, OpenAI originally targeted a $1.4 trillion valuation. The talks also include a deal for compute capacity from the Emirates. According to The Information, SoftBank's energy company SB Energy is considering delaying its own IPO because it is watching Anthropic's listing, expected in November. Source: CNBC, Financial Times, The Information.
DeepSeek: according to Bloomberg it is raising about $15 billion, two days after the US ban. The Chinese lab has raised roughly 100 billion yuan in fresh capital, with battery maker CATL and Tencent among the potential investors; it is one of the largest funding rounds in China. The money came two days after US federal civilian agencies had to remove DeepSeek technology from their systems. In June the company raised more than $7 billion at a $50 billion valuation. Source: Bloomberg, TechCrunch.
AI stocks tumbled worldwide: the Nasdaq fell 2.6 percent and Nvidia 4.4 percent. The S&P 500 had its worst day since July, even though most US stocks actually rose: the declines were driven by worries about financing data centers and chips with expensive debt. CoreWeave lost 10.1 percent and Meta 7 percent despite better-than-expected revenue, and in Asia South Korea's Kospi fell 3.5 percent. Source: The Guardian (AP).
Barclays is cutting 5,000 jobs because AI makes employees more productive. That is about 5 percent of the bank's 93,000-strong workforce. According to the finance director, about 60 percent of the cuts affect support functions, including fewer developers, and the plans should deliver more than £1.1 billion in savings in 2027. A thousand people have already been laid off this year, and the union says the bank is using AI as an excuse. Source: The Guardian, Bloomberg.
Microsoft: in a test published in Science, an AI panel of doctors diagnosed more accurately than physicians. The Microsoft Diagnostic Orchestrator has several AI agents work together to ask questions, order tests and make a diagnosis. On New England Journal of Medicine cases it reached about 83 percent accuracy, versus 36 percent for general practitioners using reference materials and 22 percent for those without; on emergency cases it reached 80 percent versus 47 percent. The system was only tested on past cases, so its real clinical impact is not yet known; the company is opening up the test suite (SDBench). Source: The Verge, Mustafa Suleyman (X).
Uber is rolling out the ChatGPT Work agent to all 30,000 employees. In an earlier 1,000-person pilot, 80 percent of users were not engineers: they used OpenAI's enterprise agent for data analysis, adapting training materials to multiple languages and markets, and building personalized, interactive dashboards. Source: TechCrunch.
Gartner: only 13 percent of companies see real returns from AI. According to a survey of 2,002 IT leaders, 58 percent of CIOs are expected to deliver AI-driven cost savings, but 51 percent say AI makes IT operations more expensive overall. Successful companies decide early which projects are worth continuing and which to stop, and plan ahead for how AI agents will reshape jobs. Source: The Decoder.
Outrage in Finland after Google cut down 250,000 trees for its data centers without a permit. According to The Guardian, at the Muhos site of its €13 billion Finnish AI infrastructure project Google cleared 330 hectares of forest without the mandatory environmental impact assessment, and about 150,000 more trees at another site. The company's head in Finland apologized, more than 93,000 people signed a petition demanding stricter data center rules, and the issue could become one of the main topics of April's election. Source: The Guardian.
Funding: $215 million for Kevin Weil's lab automation company, Mensch and LeCun back a Swiss robotics startup. Warp Sciences, the company of former OpenAI science lead Kevin Weil, raised $215 million at an $865 million valuation to make and test drug candidate molecules with AI and robots. Swiss Enact received a $30 million pre-seed (earliest-stage) investment, from backers including Arthur Mensch (Mistral), Yann LeCun and Thomas Wolf (Hugging Face). Atlanta-based Tractian, which uses sensors and AI to predict machine failures, raised $130 million at a $1.4 billion valuation, factory robot installer Formic raised $50 million, and Nirva, which makes a pendant that only listens to its wearer's voice, raised $25 million ($33 million in total). Source: TechCrunch, Tech.eu, The Information, Axios, TechCrunch.
Rumors
Kimi “Kibitz”: a mysterious new model showed up on Code Arena, probably from Moonshot. According to Chetaslua, a user who is often first to flag model leaks, a new model called “Kibitz” is running on the blind model comparison site for programming tasks; it is very fast and excellent at drawing with code, and his contacts say it will be a very efficient Kimi model. Reliability: medium. The source has been accurate several times before, but Moonshot has confirmed nothing. Source: chetaslua (X), chetaslua (X).
What this means for you: The common thread of the day is that AI agents are getting in everywhere, from your bank account to company email, so password managers, banks and cloud providers alike are working to make sure an agent can only act with approval and in a verifiable way. If you plan to launch agents in your own organization, it pays to decide at the start what they can reach and who approves their steps; our AI governance and oversight service gives you the framework for that, and for everyday tasks you can choose from our agents with built-in approval points. That way the agent speeds you up without slipping out of your control.