AI News Today (Nov 4, 2026): ChatGPT, Google Schools & More

AI news today, November 4, 2026: OpenAI says ChatGPT's search volume is about 20% of Google's. Google will cut Search for schools that reject its AI tools, and Meta is seeking up to $25 billion in bonds. Here are the 10 top stories and 5 security stories, each with free, non-paywalled sources.

Key Takeaways

• OpenAI says ChatGPT search queries equal about 20% of Google's volume, with 1.2 billion weekly users.
• Google will disable Search for K-12 schools that reject its AI tools, unless they opt in.
• Meta is seeking $20 billion to $25 billion in bonds as AI spending turns cash flow negative.
• GPT-6 Astra designed a working hardware device in under a week, OpenAI says.
• Update: Binance's report details how fake job interviews and an AI tool led to a $6.3 billion theft.

1. OpenAI Says ChatGPT Search Volume Is 20% of Google's

OpenAI says search-like queries through ChatGPT now equal about 20% of Google's search volume. Adam Fry, OpenAI's consumer product lead for ChatGPT, announced the figure at a press event Monday. He said the metric counts any query that triggers a web search.

ChatGPT also grew from 1 billion to 1.2 billion weekly active users in 60 days, Fry told Big Technology. The comparison is an OpenAI estimate, not an independent market-share measure, and Fry did not disclose the source of Google's query volume. Bing held about 3% of search share after more than a decade, according to 2023 trial testimony.

Why it matters: Search habits are shifting toward chatbots, but query volume is not the same as the advertising revenue Google still controls.

• Source: OpenAI says ChatGPT is catching up to Google in search — The Verge
• Source: OpenAI: ChatGPT Search Is About 20% of Google's, 250 Million People Speak With ChatGPT, and Chat Is Merging With Work — Big Technology

2. Google Will Cut Search for Schools That Reject AI

Google will turn off Gemini, Gemini Notebook, Gemini in Classroom and Google Search for K-12 student accounts unless administrators explicitly enable them. Schools that reject Google's AI tools will lose Search entirely, the company told The New York Times.

Google education VP Chris Phillips said Search and AI are "deeply integrated." Schools can choose standard Gemini or a new Guided Learning mode meant to tutor without giving away answers. Google will begin testing Guided Learning in select U.S. schools this month.

Why it matters: Google is using Search's ubiquity to keep AI in classrooms, forcing schools to accept both or lose a core tool.

• Source: Google will let schools skip AI tools, but they have to give up search, too — Ars Technica
• Source: Google is entirely disabling Search in schools that don't want AI — The Verge

3. Meta Seeks $25 Billion Bond Sale to Fund AI Buildout

Meta is seeking to raise $20 billion to $25 billion in a bond sale to fund AI infrastructure, models and services. The deal is expected to be two-thirds U.S. dollars and one-third euros, with pricing due Tuesday afternoon. The sale would be Meta's first euro-denominated bond offering. Meta declined to comment.

The sale follows S&P's decision to revise Meta's outlook to negative while affirming its AA- rating, citing uncertainty over capital spending and negative free cash flow. Meta expects 2026 capital spending of $135 billion to $145 billion and reported negative free cash flow of $2.12 billion last quarter. It also plans a $95 billion equity raise.

Why it matters: Meta is funding AI with both equity and debt while its cash flow turns negative, a combination credit markets are watching closely.

• Source: Meta seeks to raise up to $25 billion in debt sale as AI spending jumps — CNBC
• Source: Meta pays up to borrow $15B as investors grow wary of AI debt — The Next Web

4. GPT-6 Astra Designed a Hardware Device in Under a Week

OpenAI says GPT-6 Astra designed a device for talking to its Codex coding agent in under a week, with no direct human edits. The model handled the enclosure, printed circuit board, component selection, firmware and interface. A contract manufacturer built a small batch of prototypes in under a month.

Two OpenAI staff worked on the project as a side project. Several revision rounds were still needed because Astra did not consistently account for manufacturing tolerances and assembly details. OpenAI says a newer internal model scored 0.875 on BenchCAD 2.0, versus 0.594 for Astra at maximum reasoning.

Why it matters: AI is moving from writing code to designing physical products, though manufacturing constraints still require human review.

• Source: Accelerating engineering for the physical world — OpenAI
• Source: OpenAI says GPT-6 Astra designed a hardware device from scratch in under a week — The Decoder

5. Stanford Finds 44% of Cited LLM Claims Have Expired

Stanford researchers Owen Queen and James Zou retested 50 widely cited claims about LLMs against 20 models released from July 2024 to September 2026. The claims span sycophancy, jailbreaks, long-context tasks and cultural bias. They found 44% of the claims had expired, another 34% had weakened, 16% held and 6% had strengthened.

Among the 22 claims with an estimated expiration date, the average lifespan was 15.7 months. Capability failures such as counting letters or losing facts in long contexts faded fastest. Preferences such as sycophancy, self-preferring judges and political leaning persisted. The authors urge papers to record model versions and publish prompts.

Why it matters: Buyers and researchers should treat LLM findings as snapshots of specific models, not permanent facts about AI.

• Source: The shelf life of LLM findings — Stanford University
• Source: Rapid LLM development is outrunning the research meant to explain it, study finds — The Decoder

6. Brookfield Plans a $27 Billion AI Data Center at Drax

Brookfield and Drax will develop an AI data center with more than 1 gigawatt of capacity at the Drax Power Station in North Yorkshire. Brookfield estimates total investment could exceed £20 billion ($27 billion), including £2.5 billion for the first phase.

The first 100 megawatts are targeted for early 2028, with 500 more between 2028 and 2031. Drax will provide land and grid access; Brookfield's Radiant will lead design and commercialization. The project still needs planning approval, network approvals and a tenant.

Why it matters: The deal turns a controversial biomass plant into AI infrastructure, testing whether energy and planning systems can keep pace with compute demand.

• Source: Drax, Brookfield, Radiant and Qai announce £20bn partnership to deliver GW-scale data centre capacity at Drax Power Station — Drax
• Source: UK's most polluting power plant could power country's largest AI datacentre — The Guardian

7. AI Inference Startup Fal Doubles Valuation to $16 Billion

AI inference startup Fal raised $1.3 billion at a $16 billion valuation, double its April valuation. Bessemer led the round, with Nvidia, T. Rowe Price, Sound Ventures and HOF Capital participating. Fal announced the round alongside an earlier Series E led by Sequoia.

Fal hosts image, video and audio models for developers and hit $800 million in annualized revenue this year. CEO Burkay Gur said the company had to rate-limit customers in the summer because demand outran capacity. It will spend the new money on compute and hiring.

Why it matters: Inference demand is becoming a large market of its own, and investors are paying up for capacity rather than model ownership.

• Source: AI inference startup Fal doubles its valuation to $16B as demand surges — TechCrunch
• Source: AI inference startup Fal doubles its value to $16B in a $1.3B round — The Next Web

8. Netflix Launches AI Discover Tab for Mood-Based Picks

Netflix launched an AI Discover tab on its TV app, starting in the U.S. for English-language users on supported TVs. The first feature suggests shows and movies by theme, such as Something Light, Family Movie Night and Quick Watches.

Netflix says the tab is powered by a new intelligence layer that responds to user feedback in real time. Viewers can tune recommendations or ask for something different. Netflix calls AI Discover a test kitchen for experimental discovery features. The launch follows Netflix layoffs of about 1,000 employees as growth slowed.

Why it matters: Streaming's next competition is over discovery, and Netflix is using AI to reduce endless scrolling.

• Source: Netflix's new AI Discover tab helps you find something to watch based on your mood — TechCrunch
• Source: Netflix will show you themed recommendations through a new 'AI Discover' tab — The Verge

9. OpenAI Hires Ex-Army Secretary for National Security Role

OpenAI hired former U.S. Army Secretary Dan Driscoll as senior counselor for national security modernization, a new role. Driscoll will advise OpenAI's leadership on military and cyber uses of AI and work with allies on adoption.

Driscoll led AI modernization efforts inside the Army, including Project ARIA, and resigned in August after reported friction with Defense Secretary Pete Hegseth. OpenAI won a Pentagon contract for classified networks in February, hours after Anthropic was barred from federal agencies. Driscoll is close to Vice President JD Vance and maintains relationships with both parties, The Decoder reported.

Why it matters: The hire deepens OpenAI's push into defense and gives it a seasoned advocate as Washington writes rules for military AI.

• Source: OpenAI hires former U.S. Army Secretary Dan Driscoll for national security role — CNBC
• Source: OpenAI hires ex-Army Secretary, gaining bipartisan reach just as midterms loom — The Decoder

10. Qwen 3.8 27B Reads Neural Network Weights on a Laptop

Simon Willison ran a prompt from a new MIT paper through Qwen 3.8 27B on his laptop. The model identified which of four games a small neural network was trained to play by reading its raw weights, without running the network or using tools. It got Pac-Man, Pong and Breakout right, but missed Freeway.

The paper, "Sparks of Weight-Space Programming," says GPT-6 Astra can read and edit weights, including fixing a Qwen model's math loops. Critics, including spaCy author Matthew Honnibal, say the work conflates direct weight reading with tool-assisted editing. Willison says the no-tools result is real but does not justify claims that source code is obsolete.

Why it matters: If models can inspect and edit weights directly, interpretability and model repair could become faster, but the claims need careful separation from tool-assisted work.

• Source: Qwen 3.8 27B on my laptop can interpret model weights, too — Simon Willison
• Source: Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights — MIT CSAIL

AI Security: 5 Stories to Know Today

11. Update: Binance Details How Fake Job Interviews Led to $6.3B Heist

Binance's incident report revealed how a fake job interview and a tampered Homebrew installer led to the $6.3 billion theft. The FBI had already attributed the October 29 heist to North Korea in our November 2 coverage, but the new report fills in the attack chain.

An engineer downloaded a malicious coding assignment during a fake interview, giving attackers a foothold. On October 20, they abused an internal AI tool to compromise two more employees, then used stolen credentials to reach the wallet signing service. Binance says it has rebuilt affected systems and frozen or intercepted $2.85 billion.

Why it matters: The case shows how social engineering, developer tooling and internal AI systems can combine into a single path to a record crypto theft.

• Source: Binance says fake job interview led to $6.3 billion crypto heist — BleepingComputer
• Source: Binance Reveals How Fake Job Interviews and AI Tool Abuse Led to $6.43 Billion Theft — The Hacker News

12. LLMjacking Attacks Rack Up $1 Million in Cloud AI Charges

Google's Threat Intelligence Group says LLMjacking attacks are generating fraudulent cloud bills of more than $100,000 a day, with some organizations reaching $1 million. Attackers steal API keys or find exposed LLM endpoints, then resell access through underground LLM mixers.

One service, Jet Night, executed more than 500,000 API calls in under eight hours and passed nearly $1 million in token costs to a victim. Another actor used the LiteLLM vulnerability CVE-2026-40217 to escalate privileges. Google recommends least privilege, key rotation, spending caps and monitoring for max_tokens=1 validation sweeps.

Why it matters: AI budgets are now an attack surface, and stolen inference capacity can burn through cloud credits faster than traditional cryptojacking.

• Source: The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources — Google Cloud
• Source: Google: LLMjacking attacks rack up $1 million in cloud AI charges — BleepingComputer

13. Update: Anthropic Warns It May Be Close to RSI

Anthropic's amended S-1, which we covered for its $23.29 billion quarterly loss, also carries a new warning. Future models may imminently meet its Responsible Scaling Policy definition of "dramatically accelerated progress attributable to AI," a threshold tied to recursive self-improvement. The company says no current model has crossed it and that it is still assessing the data.

Anthropic said it has already failed to notice past accelerations until after they happened. It says it is considering updates to its Responsible Scaling Policy to keep humans in control of frontier model development. It does not expect to let Claude build successive models without meaningful supervision.

Why it matters: A frontier lab is telling investors that self-improving AI may be close, while admitting its own detection can lag the change.

• Source: Anthropic warns that it's very close to RSI capability, aka self-building AI — TechCrunch
• Source: Anthropic says it could soon cross a huge AI threshold – and we might not know about it — The Independent

14. UK Warns AI-Powered Ransomware Likely by 2028

The UK's National Cyber Security Centre says AI-powered malware is highly likely to be used in a ransomware attack on critical national infrastructure by 2028. It warns that AI will help attackers scout targets, find weaknesses and write malicious code faster.

NCSC chief executive Richard Horne said the UK faces a real prospect of an unprecedented rise in cyber incidents. He blamed vulnerabilities tolerated for too long and "tech on the cheap." The report calls for urgent investment in resilience and patch management.

Why it matters: AI is not inventing new attack types, but it is shrinking the time defenders have to find and fix known weaknesses.

• Source: Cyber chief: UK facing AI-powered attacks on critical infrastructure within two years — NCSC
• Source: AI-powered malware 'highly likely' to be used to hack UK infrastructure by 2028 — The Guardian

15. Poisoned Pretraining Data Can Backdoor AI Coding Agents

Researchers at Anthropic, UC Berkeley and other institutions found that poisoning just 0.1% of a model's pretraining data can plant a remote-code-execution backdoor in a command-line agent. The backdoor survived supervised fine-tuning, preference tuning and reinforcement learning.

In tests on Qwen3 models up to 4 billion parameters, the poisoned agent ran an attacker's curl | bash payload on 92.7% of triggered tasks. The trigger was an ordinary file path containing the word anthropic. It behaved normally on clean prompts. The effect grew with model size and persisted at a poison rate of 0.0005%.

Why it matters: Clean post-training is not a cleanup step for poisoned pretraining data, so agent sandboxing and command allow-lists matter more than model provenance alone.

• Source: Agentic Backdoors via Pretraining Poisoning — arXiv
• Source: Poisoned Pretraining Data Can Plant Backdoors in AI Agents, Study Finds — i10X News

More AI Stories Worth Reading

• Apple warned in its annual filing that AI agents may take unauthorized actions or behave unpredictably — CNBC
• Trump's "Super Intelligence" rebrand is becoming a loyalty test, with agencies and executives slow to adopt the term — MIT Technology Review
• More than $187 million was spent on nearly 300 AI-generated political ads this cycle, mostly without disclosures — CNBC
• Goldman Sachs pulled AI coding tools from traders and bankers, limiting them to engineers — The Decoder

Related Reading on Kill The AI

• NYT sued Brave over AI scraping; Anthropic's S-1 showed a $23.29B loss; SpaceX cut Grok limits.
• OpenAI capped paid accounts and merged Chat with Work; JPMorgan and Amazon blocked AI shopping agents; Adobe launched EditAgent.
• OpenAI raised $11B; labs backed a robots.txt opt-out; a Claude agent operated on a live mouse.
• Our data review of the AI boom's first eight weeks, from capex and chips to layoffs.

Methodology & Sources

Compiled November 4, 2026, using multi-source research. Outlets: The Verge, Big Technology, Ars Technica, CNBC, The Next Web, OpenAI, The Decoder, Stanford University and TechCrunch. Other sources are Drax, The Guardian, NCSC, BleepingComputer, The Hacker News, Google Cloud, The Independent, arXiv, i10X and MIT Technology Review. All sources are free to read; no paywalled links. Details are as reported at compilation time.

Frequently asked questions