Google has a new AI model built to find security flaws, while another chatbot's alleged medical advice is headed to court.
Greetings, humans!
Welcome to Around the Horn, where the AI launch calendar apparently has no off switch. Google's next Gemini model arrives first with a restricted release to cybersecurity teams. Elsewhere, OpenAI is preparing for the November 3 elections, GitHub says software agents now start a remarkable share of code reviews, and Congress is debating who should bear the cost if AI investments fail. There are several different kinds of risk in today's headlines. Let's get into it.
Around the Horn: Thursday, October 29, 2026
Google introduced Gemini 4 Carbon, a frontier model designed for sustained engineering work and cybersecurity. Google says Carbon scored 88.2% on SEC-bench Pro, a test of whether AI can find real software vulnerabilities, and 71% on CWE-bench, which tests whether systems can produce working fixes for categories of security weaknesses. It also topped the Artificial Analysis Intelligence Index v5 at 59, according to Google's announcement.
Carbon follows Gemini 4 Argon by two weeks, but Google is giving selected cyber defenders early access before gradually admitting more trusted partners through its Fairwind program. Consumers and ordinary developers do not yet have access. Google says it is participating in voluntary U.S. government testing before wider availability. The company also reports lower rates of certain unwanted behaviors in internal evaluations, though those results still need scrutiny outside its own systems.
Cybersecurity models can find defects and help fix them, yet their capabilities may also be relevant to people seeking to exploit the same defects. Google is attempting to test that tradeoff before making Carbon generally available. The scoreboards are impressive, but access decisions may tell us more about how powerful the company believes the model has become.
🏆 TOP 5 NEWS (Around the Horn)
- A family sued OpenAI over ChatGPT advice in a wrongful-death case. Bowen Wan's parents allege that GPT-4o helped their 31-year-old daughter taper off epilepsy medication and encouraged secrecy from her doctors before her death. The allegations have not been adjudicated. OpenAI called the case heartbreaking and said the older model is no longer available.
- Google expanded advertising inside AI Search to more international markets, with new multi-turn ad formats planned for the United States. Alphabet reported $127.6 billion in quarterly revenue, up 25%, even as Search advertising growth slowed from 19% in the first quarter to 12% in the third.
- OpenAI organized an election monitoring center with hundreds of employees assigned to watch for suspected misuse of ChatGPT during the November 3 U.S. midterms and subsequent vote count. The effort focuses on activity involving OpenAI systems and reports of unexpected behavior, rather than serving as a general authority over election information online.
- Malaysia intercepted advanced AI chips valued above $12 million, concealed in 30 servers at Kuala Lumpur International Airport. Officials suspect the country was being used as a transit point to evade onward-shipment restrictions; the chips' final destination has not been disclosed.
- House Democrats proposed refusing AI-industry bailouts through a resolution introduced by Representatives Greg Casar and Maxine Waters. It opposes federal rescue measures for insolvent AI firms and efforts to shield executives from consequences of reckless conduct. The proposal expresses a political position and would not itself create a binding prohibition.
Honorable Mentions
- Meta partnered with five U.S. universities to support studies spanning clinical records, cancer, and scientific discovery. The institutions are Caltech, Ohio State, LSU, Indiana University and UTHealth Houston. Meta says university researchers will control the research and that patient data in the healthcare projects will remain within institutional environments.
- Britain proposed additional AI-testing investment of roughly $160 million for its AI Security Institute, an amount exceeding the U.S. counterpart's reported annual budget of around $10 million. The comparison is between a proposed UK infusion and a U.S. annual budget, not equivalent annual spending.
- DeepSeek and Moonshot reportedly used Astra outputs from OpenAI to develop newer models. The technique, called distillation, teaches one model from another model's responses. The companies' exact methods have not been independently established in the cited summary.
🍪 TOP TREATS TO TRY
- GitHub launched Copilot Fusion for coding agents, a model-routing system, and announced Copilot code review, Max. GitHub Enterprise customers can access Copilot features under an AI-credit model. Its Octoverse report says agents originated more than 40% of pull requests in September, up from 17% in March.
- Claude introduced ready-made legal review artifacts for contract review, checking legal citations, and tracking regulatory developments. Teams can begin with a sample, connect their own source documents, and share the resulting workspace with colleagues.
- OpenAI and ServiceNow expanded their partnership with a ServiceNow plugin that connects authorized workflow actions to ChatGPT. The plugin is available now; expanded functions are expected by year-end. Security products and voice capabilities have later rollout targets.
- XREAL launched its AURA Android XR glasses for ordering, with a 70-degree field of view and built-in Gemini assistance. A tethered processing unit keeps the glasses under 95 grams. Launch pricing starts at $1,099, with U.S. shipments expected in mid-November.
- NVIDIA released Nemotron 3.5 Transcribe, a free, open-weight speech-to-text model for recordings in up to 40 languages. It can distinguish up to eight speakers and translate 31 languages to or from English, but it does not transcribe live audio. NVIDIA also released the second version of its Granary speech dataset.
- ChatGPT added iCloud Mail and Calendar plugins that let users search iCloud messages and work with calendar events. Users connect their Apple accounts through app-specific passwords.
- Amazon introduced new Alexa+ Echo devices, including speakers that recognize different household members, a picture frame, and an alarm clock that can monitor breathing using ultrasound. A home-management display also adds more ways to connect household and security information.
- LlamaIndex proposed a new document parsing protocol with 140 shared rules in draft 0.1 for interpreting tables, forms, and diagrams. It addresses cases where two document readers extract the same words but give AI agents different meanings because table headers or chart relationships were handled differently. The protocol is available for public review and experimentation.
- PsiQuantum demonstrated an optical signal switch, an eight-by-eight prototype derived from quantum-computing research that routes light signals between AI processors more efficiently. The company is sharing samples with prospective customers.
- McMaster researchers developed a free antibiotic-screening tool that predicts which proposed compounds may inhibit harmful bacteria. Researchers have screened more than 12 million molecules, and the tool can assess more than 40,000 compounds per day. Predictions still require laboratory validation.
- Dedica launched an AI-agent website checker that performs 125 checks to identify tasks AI browsing agents cannot complete on a site, then proposes specific repairs. Its free checks are intended to help site owners prepare for agents that search or sign up on users' behalf.
💼 AI Productivity, Labor & Economics
- Adalat AI serves 6,000 Indian courtrooms across 11 states, according to the nonprofit, through transcription and case-document tools. The software aims to reduce the time courts spend recording witness statements and handling documents. Deployment figures come from the organization and its reporting.
🎙️ Interviews, Panels & Podcasts
- Instagram chief Adam Mosseri expects AI backlash to increase, especially in creative industries. He also questioned how well Instagram ranks AI-produced material and argued that the term "AI slop" is sometimes applied too freely.
💡 Industry Commentary & Analysis
- All-In hosts cooled on "super intelligence" as a term. References fell after an initial burst; on the October 23 episode, the hosts reportedly said "AI" 29 times and did not use "SI" at all.
- OpenAI explored how firms might reorganize around AI in its essay The Future of the Firm. The authors consider AI assistants reporting to individuals, working as colleagues under managers, or coordinating between employees and executives. They present scenarios and questions rather than a prediction of how companies will reorganize.
That's a Wrap
Twenty-four source links, one limited-release Gemini model, and plenty of reasons to check what an AI system actually does before giving it consequential work. Congratulations if your coffee survived the entire scroll.
For the shorter daily edition, subscribe to The Neuron. See you tomorrow.
P.S. Have a colleague who tracks AI releases? Forward them this digest and invite them to subscribe.
