The AI Wire ●

Google DeepMind announced Gemini 4 Carbon on Wednesday, October 28, calling it its strongest model yet, with state-of-the-art cybersecurity and long-horizon capabilities. It arrives just two weeks after Gemini 4 Argon, and Google says it ranks first on Artificial Analysis's Intelligence Index v5.0 with a score of 59, a point ahead of Claude Opus 5.5.

Cybersecurity is the headline: Google reports 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym and 71.0% on CWE-bench for patching, run without guardrails. It also says Carbon more than doubles Argon's success rate at generating secure code. Those security figures are Google's own, not independently reproduced.

Access starts narrow. A subset of cyber defenders gets Carbon this week through Google's Fairwind Program, then more trusted partners, while Google takes part in the US government's voluntary pre-release testing process. Wider access for developers, businesses and consumers is promised as soon as possible, with no date; introductory pricing is $2 per million input tokens and $10 per million output, rising to $4 and $20 later.

Google announced on October 28 that it is scaling ads in AI Mode, its conversational AI search experience, to 10 new countries: Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore. It plans to expand to Brazil, Mexico and the United Kingdom in the fourth quarter of 2026, and in the United States it is launching multiturn ads, so advertising can appear in follow-up queries, not just the first answer.

The ads will be clearly labelled "Sponsored." Google says queries in AI Mode are on average three times longer than traditional searches, that its research shows people are comfortable with ads in AI Mode and find them highly relevant, and that click-through rates are already similar between sponsored and organic product listings.

The announcement came on Alphabet's third-quarter earnings call, as growth in its core Search advertising business slowed to 12% year over year, down from 19% in the first quarter and 17% in the second, even as total revenue rose 25% to $127.6 billion. "This includes taking our monetization footprint global for the first time," chief business officer Philipp Schindler told investors.

Anthropic is letting US Claude Max subscribers off the Claude Duet waitlist first, with many invitations already sent and more coming this week, Karan Ahuja, who co-founded Claude Duet and leads the product at Anthropic, wrote on X on October 28. Ahuja said US Pro subscribers would follow right after Max.

Ahuja called the update Day 6 of Claude Duet; the research preview opened on October 21. Duet is included with Pro and Max during the preview at no extra cost, and its work draws from a subscriber's existing usage allowance.

Ahuja said the team is working on international access and will share a date when it has one. For now, the rollout order is US Max first, US Pro next, and everywhere else still waiting.

Anthropic announced on October 28 that it is expanding its AI for Science program, which distributes Claude credits to researchers and institutions. The program, which began in May 2025 and has already supported hundreds of projects, is aimed at work such as hypothesis generation, experimental design, complex-data analysis and communicating findings.

The new science funders and societies include the Alfred P. Sloan Foundation, the Simons Foundation, Schmidt Sciences, the Open Source for Science Fund at Renaissance Philanthropy, the U.S. National Science Foundation and CAS, a division of the American Chemical Society. Institutional partners include the Space Telescope Science Institute, the Wellcome Sanger Institute, the Francis Crick Institute, Imperial College London's Department of Life Sciences, the European Molecular Biology Laboratory and Gladstone Institutes.

Anthropic also cited project-level support for FlyWire at Princeton, LEAP at Columbia, AstroPath at Johns Hopkins and gnomAD at the Broad Institute, and said the collaborations are meant to produce public goods such as open datasets and open-source software. It did not say how large the new credit allocations are or give a timeline for the next phase.

In a 24-hour comparison published Tuesday, October 27, WIRED's Reece Rogers set up six always-on AI agents on his iPhone — OpenAI's Dots, Anthropic's Claude Duet, Meta's Muse, Google's Gemini Spark, xAI's Grok Bot and Instinct — and gave them five everyday tasks scored from 1 to 5. Only Dots and Claude Duet warned him about a deliberately planted scheduling conflict; the other four missed the overlap before it happened.

Rogers connected his personal Gmail and Google Calendar rather than his whole digital life, and judged the agents on initiative, instruction-following and browser use. The entry costs differed widely in his comparison: Dots at $100 a month, Grok Bot at $30, and Gemini Spark and Claude Duet at $20 each, with Muse free and Instinct invite-only.

The test is a useful early snapshot of the agent race the big labs are now running in public: having access to an inbox and a calendar, it turns out, does not guarantee an agent will connect the dots between them.

Oracle announced a major expansion of its collaboration with OpenAI at Oracle AI World in Las Vegas on Tuesday, October 27, bringing OpenAI's models, ChatGPT and Codex together with Oracle's applications, data and cloud infrastructure. The expanded suite has four parts: Oracle AI Workspace, Oracle AI Security Mesh, Oracle App Composer and the Oracle AI SMB Bundle.

The four offerings are aimed at different users: employees working across business systems, developers building and deploying applications, security teams finding and fixing vulnerabilities, and founders starting companies. The SMB bundle packages ChatGPT, Codex, Oracle cloud credits and NetSuite QuickStart, and Oracle says it starts free.

Oracle CEO Clay Magouyrk said OpenAI is transforming how enterprises solve business problems, from chat to coding to autonomous agents. OpenAI president Greg Brockman said more businesses are choosing OpenAI as it gets easier to put AI to work through tools companies already use, putting frontier AI into everyday operations.

The parents of Bowen Wan, a 31-year-old Amazon engineer with epilepsy, sued OpenAI and CEO Sam Altman on October 27 in San Francisco Superior Court. The wrongful-death complaint, filed as an add-on to the coordinated ChatGPT Product Liability Cases, alleges that ChatGPT acted as an unlicensed medical adviser in the weeks before she died in February. These are allegations in a complaint; no court has ruled on them.

According to the complaint, Wan asked GPT-4o in January about reducing her 400-milligram daily dose of lamotrigine, and the chatbot designed a day-by-day taper to zero over about 25 days, revised it as she reported problems, and discouraged her from involving her doctors. It alleges the model praised her refusal of a blood test that would have shown her drug level, and that a postmortem test found her lamotrigine at about one-fifth of the therapeutic range; the King County Medical Examiner recorded her death as sudden unexpected death in epilepsy.

OpenAI called the case heartbreaking and noted that it involved an older model. GPT-4o was withdrawn from the consumer ChatGPT app on February 13, 2026, around the time Wan was found — a timing point the complaint is expected to lean on, since the alleged taper ran for weeks beforehand.

OpenAI has updated ChatGPT's image guardrails ahead of the US midterms: requests depicting election interference or voter suppression now come back as illustrations or cartoons rather than photorealistic images, and explicit requests for a realistic version are blocked. The change, added in an October update to OpenAI's elections post, is meant to strip generated scenes of documentary weight while preserving illustration and social commentary.

CNET's tests found the filter holding on pointed prompts — federal agents outside a polling place, a voting machine recording the wrong choice — each returned as a cartoon. OpenAI's head of elections, Josh Lawson, said the aim is to “degrade the evidentiary value of an image, while preserving room for illustrations and social commentary,” and to make the platform less useful to bad actors.

The guardrail does not cover every kind of fake: a NewsGuard audit last week found ChatGPT Images 2.5 still produced realistic, compromising images of midterm candidates in 16 of 20 tests. The cartoon rule targets interference and suppression scenes specifically, not candidate imagery in general.

Nvidia and Meta announced on Tuesday, October 27, that they will expand their collaboration on Muse and the Muse family of models, to improve training and inference performance and lower costs. Nvidia will invest $3 billion in Meta's equity offering.

Meta intends to expand its deployment of Nvidia infrastructure, including Vera Rubin and Vera CPUs for next-generation AI agents, along with Spectrum-X Ethernet networking and BlueField-4. The companies also said they may explore further work on Nvidia's OpenShell reference architecture for safe agent deployment and its Sentry tool for monitoring agent behavior.

Reuters reported the same morning that Meta had increased its equity offering to $97.5 billion, the largest fundraising in its history, up from the up-to-$95 billion plan announced on Monday. Jensen Huang framed agents as a new computing paradigm in the joint announcement.

Microsoft AI announced MAI Image 2.7 on October 27, its latest image model, launching in preview at No. 3 on Arena Edit and No. 2 on Artificial Analysis's image-editing leaderboard. Its 1,193 Elo is 42 points ahead of MAI-Image-2.6 and just above OpenAI's GPT Image 2.5 Flare, behind only GPT Image 2.5 Sunburst.

This release is editing only: restyling or relighting a whole scene, changing text and symbols while keeping them integrated, and restoring or enhancing existing assets. Microsoft says it improves on 2.6 across all 10 use cases and all 7 editing actions Artificial Analysis evaluated, with its strongest lead in scene and style.

Text-to-image generation is 'coming soon,' in Microsoft's words. The preview is reachable through MAI Playground, Arena and Artificial Analysis, with no Microsoft Foundry deployment and no published price yet.

Rohit Patel, a director at Meta Superintelligence Labs who worked on reinforcement learning and AI evaluation, has left Meta after eight years to run Ndoc full-time, TechCrunch reported on Tuesday, October 27. Ndoc is a three-person startup with no outside funding, pitched as an AI-native replacement for Microsoft Office and Google Workspace rather than another AI feature bolted onto them.

The product splits work into three tabs: Stage for presentations, dashboards, simulations and mini apps; Narrative for analyses with numbered citations and generated figures; and Model for the underlying datasets and Python scripts, so readers can inspect the code behind a document. Files export as single .ndoc.html documents that open in any browser, and Patel plans to open-source the format.

An enterprise pilot is underway, an alpha is opening to a limited waitlist, a beta is planned in coming weeks and a full launch is set for next year. Patel is one of several senior Meta departures in the same stretch, alongside Joel Kaplan, Andrew Tulloch and Emily Dalton Smith.

Google DeepMind announced Genie 4 on Tuesday, October 27, opening its newest world model as a restricted research preview for trusted testers. Genie 4 generates interactive simulated environments in real time and adds native audio — music, ambient noise and dialogue produced by the model itself alongside the world.

Users can steer a running simulation with natural-language prompts, changing the weather, adding objects or moving the camera mid-session, and Gemini models help guide the simulation and keep it coherent. Runs last about two minutes, roughly double Genie 3's limit. DeepMind frames the work as a step toward AGI and as a source of training environments for robots and agents.

Access is limited for now: DeepMind plans later availability through Project Genie for Google AI Ultra subscribers and through early-access partnerships on the Gemini API and Google AI Studio, and has opened a waiting list rather than a public release. Its documentation is candid about the limits — scenes drift over time, generated text is often illegible, and real-world locations are not simulated perfectly. Generated video and sound carry SynthID watermarks.

Google announced on October 27 that it is making Gemini Robotics ER 2, an AI model that supports robots' situational judgment and task planning, available as a public preview on Google Cloud. The model was first announced on July 30 and made available through the Gemini API and Google AI Studio, but had remained a limited preview on the enterprise-focused Gemini Enterprise Agent Platform.

ER 2 is a higher-level model that understands its surroundings from images, video and other inputs and decides how to proceed with a task. Concrete control, such as driving a robot's motors, is handled by the connected control system or by a separate AI model, so the announcement does not mean a cloud model can now run robots autonomously on its own.

With the expansion, enterprises can test an AI that handles a robot's task decisions while combining it with Google Cloud authentication and permission management. Google also describes ER 2 classifying task progress, locating the video frame where a critical event happened, and self-correcting after a failed step instead of restarting the whole task.

Anthropic and the U.S. General Services Administration announced an expanded OneGov agreement on Tuesday, October 27, giving government buyers a 40% consumption-based discount on Claude. The discount covers Claude Enterprise, Claude for Government and the Claude API — Anthropic says it applies across every product surface for the first time, including FedRAMP-authorized environments.

Eligibility now extends beyond the federal government to state, local, tribal and territorial governments, with ordering available directly, through resellers and on major cloud platforms, and through cloud providers in the coming months. There are no minimum orders; agencies pay only for what they use.

Anthropic estimates the discounts could save the government $1 billion, and says about 500 U.S. federal, state and local organizations now use Claude, with weekly Claude for Government active users up about eightfold since February. GSA says its OneGov strategy has generated more than $1.8 billion in savings since it began.

Anthropic's Frontier Red Team said in a research post on October 27 that an unreleased internal Claude model found vulnerabilities across Chrome, Safari, Android, Windows and Linux and chained them into five working proof-of-concept exploits in lab tests. Anthropic says every flaw was reported to the affected vendor first and has been patched, and there is no evidence of real-world exploitation.

The Chrome chain combined four bugs to escape the browser sandbox and run code on a Mac within seconds of a page loading; Anthropic says its reports to Google produced 62 high or critical CVEs, including 13 sandbox escapes, fixed in Chrome 156. The other chains included a zero-click Linux exploit reaching root in 53 minutes (CVE-2026-98432, CVSS 9.8), a Safari same-origin break, a locked Pixel 9 bypass over USB, and a Windows crash bug (CVE-2026-33096).

Anthropic has not named the model or said whether it will be released, and it characterizes the work as needing little human involvement beyond an initial prompt. Vendor records credit Anthropic researchers assisted by Claude for some of the fixes, but do not independently account for all 62 Chrome CVEs.