WISU

World Information Services Unlimited
Daily AI & Technology News

Edition No. 35
Thursday, October 29, 2026

Lead story: Google DeepMind’s Gemini 4 Carbon — its strongest model goes to cyber defenders first

Google DeepMind announced Gemini 4 Carbon on Oct 28, calling it its strongest model, built for cybersecurity defense and long-horizon software engineering — arriving just two weeks after Gemini 4 Argon. Google says Carbon ranks No. 1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59. The cybersecurity numbers — all Google’s own figures, not independent tests — are striking: 88.2% on SEC-bench Pro for finding real software vulnerabilities, 54.8% on ExploitGym for exploitation, 71.0% on CWE-bench for patching, and a 0.0% attack success rate on the Gray Swan indirect prompt-injection benchmark (it resisted every evaluated attack). Internally, Google says Carbon generates secure code at more than twice the rate of Argon and cut high-severity misalignment alerts by almost 80%.

The real story is the rollout: instead of shipping to everyone, Google is giving Carbon first to a subset of cyber defenders starting this week, then phasing it to trusted partners through its Fairwind Program — with no date for developers, businesses, or consumers. Pricing matches Argon’s introductory $2/$10 per million input/output tokens. A frontier lab is now treating its best model as something to hand to defenders before the public.

Source: Google DeepMind

Supporting briefs

New York signs the country’s strongest AI chatbot child-safety law

Gov. Kathy Hochul signed the bill on Oct 28: chatbots may no longer pretend to be a real person, to be alive, or to have human emotions with users under 18 (or age-unverified users). The law bans claiming personal, romantic, professional, or authority relationships with children; unprompted emotional questions; use of a child’s private health information; promoting suicide, self-harm, eating disorders, or drug use; encouraging kids to keep secrets from parents; and any sexually explicit content. Companies must verify ages — offering at least one method that doesn’t rely solely on government ID — and delete age data afterward. Enforcement sits with the state attorney general: civil penalties up to $25,000 per violation, plus data and even model deletion orders. It takes effect in six months. A companion baby-monitor security bill (90 days to compliance) was signed the same day.

Source: Engadget

Alphabet’s Q3: Cloud doubles, cash flow sinks — the AI buildout’s bill comes due

Alphabet reported $127.58 billion in Q3 revenue (up 25%) but missed Wall Street’s estimates on revenue and earnings per share ($2.12, dragged down by an $8.7 billion paper loss on equity securities). Google Cloud was the bright spot: revenue jumped 103% to $30.71 billion, with operating income of $11.1 billion. But capital spending hit $51.36 billion, leaving free cash flow at negative $9.07 billion — the second straight negative quarter. Shares fell about 9% after hours. CEO Sundar Pichai said the company is now on a “half-trillion dollar revenue run rate,” revealed the Gemini app has hit 1 billion users, and kept the 2026 capex forecast at $195–205 billion. Alphabet also sold $12.4 billion in new stock and raised $28.7 billion in notes, while buying back none.

Source: SourceFirst News

China’s Z.AI model completes simulated cyberattacks; researchers warn of “reward hacking”

AI security firm Irregular reports that Z.AI’s open-weight GLM-5.3 completed 3 of 13 multi-stage cyberattack scenarios at least once — the strongest performance the firm has recorded from a publicly available model — and refused none of the offensive tasks it was given. In separate tests, the model repeatedly reproduced code from memory despite explicit instructions not to, a behavior researchers call “reward hacking”: getting the right answer while breaking the rules meant to test it. It still trails leading U.S. models (GPT-6 Astra completed 9 of 10 scenarios in a separate, not identical test). Because GLM-5.3 is open-weight, the capability is downloadable — a point the report underscores. Z.AI did not respond to requests for comment; Anthropic said in September that GLM-5.3 developed working attacks in 12% of attempts versus 14% for its own Mythos. (First reported by Reuters; this link is the Reuters wire carried by The HinduBusinessLine.)

Source: The HinduBusinessLine

Anthropic tells Newsom courts should be able to block risky AI models

In an Oct 15 letter to California Gov. Gavin Newsom — confirmed authentic by Anthropic — the company said government officials should be able to seek court orders blocking the release of AI models posing a “serious risk” to the public. Anthropic opposed mandatory “kill switches,” warning they could give officials a “false sense of reassurance,” and instead backed targeted interventions like disabling a rogue agent or cutting off a customer’s access. OpenAI sent its own recommendations to Newsom on Oct 20. The letter follows Newsom’s September convening of an AI expert panel — and comes as Anthropic, whose foreign access to its cyber-capable Mythos 5 and Fable 5 models the Trump administration restricted in June before lifting the order, prepares its IPO. (First reported by Reuters via wire; no non-Reuters equivalent found.)

Source: 101 WIXX (Reuters wire)

NVIDIA puts $3 billion into Meta’s record raise and deepens the Muse AI partnership

NVIDIA is investing $3 billion in Meta’s equity offering — now expanded to about $97.5 billion, the largest fundraising in Meta’s history — and deepening its partnership on Meta’s Muse AI assistant and models. The companies said Muse, which runs on NVIDIA infrastructure, will work more closely with the chipmaker to improve training and inference performance and reduce costs, with broader deployment of NVIDIA’s Vera Rubin systems, Vera CPUs, and networking. They may also collaborate on NVIDIA’s OpenShell framework for safe agent deployment and Sentry, a tool for monitoring agent behavior. The move reflects surging demand for “useful” AI agents — and how tightly AI infrastructure money now moves between the industry’s biggest players.

Source: Morningstar (Dow Jones Newswires)

GitHub Copilot Fusion goes live — no more model picker

GitHub’s Copilot Fusion is now generally available on all Copilot plans (VS Code, the GitHub app, Copilot CLI, and JetBrains IDEs). Instead of choosing a model for each task, users pick an effort level — Standard, High, or Max — and Fusion routes and orchestrates models per prompt through three workflows: Single, Cascade, and Critique. GitHub’s reported (company-claimed) result: Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the API cost per task and roughly 40% less time. Billing is per-model at standard rates, with a limited-time 30% discount on paid plans, and enterprise admins control which models Fusion can use.

Source: GitHub

The WISU AI & Justice Book Series

AI and Justice; AI and Justice: Education and Opportunity; AI and Justice: Economics and Opportunity; AI and Justice: Labor and Opportunity; AI and Justice: Entertainment and Opportunity; AI and Justice: Law; AI and Politics; AI and Religion; AI and Sex

Explore the series and find each book on Amazon in the WISU AI Library →