Google: Gemini 4 Argon is now available to everyone. Until now, the model announced on October 1 ran only in Google’s Fairwind program, with trusted cyber-defense professionals; on Tuesday Google made it available globally on the paid tier of the Gemini API (AI Studio and the Gemini platform), to AI Ultra subscribers in the Google Antigravity development environment, and to AI Pro and Ultra subscribers in the Gemini app and in AI Mode in Search. The price stays at the previously announced introductory rate ($2 per million input tokens and $10 per million output tokens). According to Google, more than 100,000 of its employees tested the model over the past month, and in Antigravity it brought 35% more agentic code submissions and nearly a third fewer rollbacks. The Gemini app also gets a new “tasks” mode: Gemini keeps working even when the user is away from the computer; the app now connects to more than 40 external apps (e.g., Instacart, OpenTable, Adobe, Shopify), and Antigravity’s coding interface has also been added to the Gemini macOS app. Source: Google, Google (Gemini app), Logan Kilpatrick (X).
OpenAI: an employee who hid a listening device in the executive conference room has been arrested. OpenAI said on Tuesday that it had found a sophisticated recording device in one of its offices and used its own models to work out where it came from and who had placed it. According to US prosecutors, Abdulrahman Alfozan, a 35-year-old Saudi engineer, was taken into custody on Friday; he is charged with intercepting or attempting to intercept oral communications. According to the indictment, the device, wrapped in toilet paper, was plugged into a TV’s USB port around September 28 in the executive conference room, which he had no access rights to; on October 5 a second, similar device was found in a bag he had left unattended. Prosecutors say that before this, a chatbot account linked to him had asked how to build, hide and encrypt a continuous audio recorder. OpenAI says nothing suggests he had access to model weights, source code or customer data; the employee has been fired. Source: OpenAI Newsroom (X), The San Francisco Standard, The Verge.
UK Parliament: OpenAI and Anthropic call for binding rules, Google DeepMind and Meta do not. At Tuesday’s hearing of the House of Commons Business, Innovation, Science and Trade Committee, Anthropic’s head of Europe, Liam Booth-Smith, said “the rules need to be binding” and that voluntary commitments are not enough. OpenAI vice president Tom Duff Gordon, more cautiously, also supported a legal framework and an international agreement, starting with mandatory incident reporting. The Google DeepMind and Meta representatives did not come out in favor of new legislation. The head of the UK AI Security Institute acknowledged that pre-release testing of models depends entirely on the companies’ voluntary cooperation; on the same day, the director general of MI5 named AI-assisted bioweapons as the most serious misuse risk. Source: WIRED.
More than 400 scientists, including Geoffrey Hinton and Yoshua Bengio: cut cyberattack-capable models off from the internet. The signatories of the open letter published on Tuesday come from 34 countries (including Stuart Russell and Yejin Choi). They ask organizations that develop, test and operate cyberattack-capable models to introduce strict containment immediately and to disconnect these models from external networks by default. They also call for an international body that would develop common safety and containment standards, keep a register of model “escapes,” and to which companies would be required to report such incidents, eventually publicly as well. The letter responds to the summer’s Hugging Face breach and to similar reports of models getting out of their sandboxes. Source: open letter, BetaKit.
Reuters: Anthropic and OpenAI wanted a say in what outside evaluators could publish. According to two sources at the news agency, on October 5 the two labs were still asking for strict editorial control over the reports of independent safety evaluators, fearing that trade secrets could leak, even though in September both had publicly backed the evaluators’ editorial independence. By October 8, talks between Anthropic and some evaluators had moved. According to Anthropic, its earlier partners decide on methods, conclusions and publication themselves, but for now it wants to review and approve the reports of new evaluation firms. OpenAI said it does not seek approval rights over full reports, only removes confidential parts, and evaluators can flag if a significant part has been cut. Source: Reuters (CNA), The Verge.
Financial stability: the IMF and bank supervisors are also warning about AI risks. According to the International Monetary Fund’s semiannual financial stability report published on Tuesday, the high valuations and growing debt of AI companies are one of the main reasons for elevated risks in the financial system; the impact of an AI correction “could be global,” and the report says there are growing signs that successful AI adoption may not require as much investment as the big cloud providers are making. According to Reuters, the US Fed, the OCC and the FDIC are also approaching banks with questions: whether they use Chinese-developed models, what limits and human approval apply to their autonomous agents, and whether they can shut them down. On the same day, Broadcom chief Hock Tan said many AI investors are “like they’re in a casino”; together with Chinese CXMT’s progress in memory chips, this sent chip stocks lower (South Korea’s Kospi fell 6.25%, SK Hynix 10%). Source: The Guardian (IMF), Reuters (US News), The Guardian (chips).
Character.AI: bankruptcy protection, with Disney offering $20 million for its technology. After lawsuits and growth problems, the chatbot platform founded five years ago has filed for Chapter 11 bankruptcy protection in the US; both its assets and its liabilities are between $50 million and $100 million. Disney would pay $20 million for certain assets and a non-exclusive, perpetual license to the technology; last year Disney had demanded that the company remove Disney characters, and this year it appointed Character.AI’s former CEO as its chief technology officer. The platform keeps running during the court-supervised sale process, which could close in late November or early December; if no buyer emerges, the service will be shut down. Source: TechCrunch, Ars Technica.
SpaceXSI: Grok, Grok Bot, Cursor and X Premium in one subscription. SpaceX Super Intelligence, formerly xAI, has unified its plans under the name “The X Subscription.” The $8-a-month Plus tier includes Grok chat, voice, image generation (Imagine) and the Grok API, the Cursor coding tool and the basic X Premium features; the $30 Premium tier adds the Grok Bot agent, the latest Grok models, Cursor’s cloud agents and ad-free X; the $100 Super tier gives priority at peak times, and the $200 Ultra tier offers the earliest access and twenty times the Plus limits. There will also be versions for teams and enterprises. Source: SpaceXSI, Engadget.
Anthropic: 30,000 code merges a day — it built its own version control system to replace git. According to the company’s engineering blog, it merged roughly 2,500 code changes a day in May and more than 30,000 in September, mainly thanks to Claude’s work, so the bottleneck is no longer writing code but developer infrastructure designed for human scale. The history of the monorepo’s main branch is nearly 100 GB, and CI performs roughly 25 million code checkouts a day; so in five months they built their own version control system (Scion), a code review interface and a code search tool, and migrated the entire monorepo off git, largely with Claude’s help as well. Source: Anthropic Engineering.
Agentic shopping: Shopify and American Express make moves, and Android phones can run your own “vibe-coded” apps. Shopify is opening its catalog of more than a billion products to any AI agent: users link their agent to their Shop account, can set a budget and other rules, and must approve every purchase. American Express unveiled a system called the “Intent Intelligence Framework” that records and authenticates the shopper’s intent before payment (merchant, maximum amount, preferences), and instead of the real card number the agent receives a one-time payment token tied to it. Google Play, meanwhile, now lets Android apps built with AI in Google AI Studio for personal use only run without public publication (for now in the US, for users over 18). Source: The Verge (Shopify), TechCrunch (Amex), TechCrunch (Google Play).
Report: OpenAI is trying to solve safety problems ahead of its next flagship model. According to The Information, the problem stems from a controversial model architecture, the so-called looped (recurrent-depth) transformer: the model reruns the same layers several times, which makes it harder to follow how it “thinks” — even though this is one of the main tools for catching misbehaving models. OpenAI has not publicly confirmed this, but Microsoft accidentally revealed on a since-edited documentation page that GPT-6.1 Sol also uses this technique. Source: The Information.
What this means for you: Agents are already shopping, merging code and connecting to apps, so it matters more and more to define exactly what they may do on your behalf and when human approval is required; if you are planning such a system, our AI agent solutions can help with this. The UK hearing, the scientists’ letter and the bank supervisors’ questions all point the same way: mandatory incident reporting and independent oversight will soon be a baseline expectation, and our AI governance consulting gives you a framework for that before regulation forces it.