AI News Today (Oct 28, 2026): Genie 4, OpenAI & Meta

AI news today, October 28, 2026: Google DeepMind's Genie 4 adds native audio and real-time steering. OpenAI says its models refuted a key computing conjecture, and Nvidia and AMD put $4 billion into Meta. Here are the 10 top stories and 5 security stories, each with free, non-paywalled sources.

Key Takeaways

• Google DeepMind released Genie 4, a world model that generates audio and lets users steer scenes in real time.
• OpenAI says its models refuted the Exponential Time Hypothesis; Anthropic released its own refutation hours later.
• Nvidia and AMD will invest $4 billion in Meta after Meta announced a $95 billion stock offering.
• The UK Home Office may face a watchdog probe over secret AI use in asylum decisions.
• CyberStrikeAI used 44 AI agents to probe at least 19 organizations, mostly in Japan.

1. Google DeepMind's Genie 4 Adds Audio and Real-Time Steering

Google DeepMind released Genie 4 in research preview on Tuesday. The world model generates photorealistic, interactive scenes with native audio, including ambient sound, music and character dialogue. Users can change the weather, move the camera, or add people and objects with text prompts while the simulation runs.

Gemini 3.8 Flash acts as a guiding agent, planning steps and checking frames to keep each world coherent. DeepMind says Genie 4 can hold a scene together for up to 2 minutes, double Genie 3's limit, with better physics and more natural human movement.

Genie 4 remains an early research model. DeepMind says object permanence, audio and legible text can still drift in longer sessions. Access starts with trusted testers, then Google AI Ultra subscribers and early partners in the Gemini API and AI Studio. All generated video and audio carry SynthID watermarks.

Why it matters: World models are becoming training grounds for robots and agents, and Google is trying to set the standard before rivals consolidate.

• Source: Genie 4: A new generation of world simulation — Google DeepMind
• Source: DeepMind unveils Genie 4 world model promising more realistic physics and humans — Ars Technica

2. OpenAI Says Its Models Refuted a Key Computing Conjecture

OpenAI said its models produced a proof that the Exponential Time Hypothesis (ETH) is false, one of three results it announced in math, computer science and physics. ETH, proposed in 1999, says certain logic problems require exponential time. OpenAI published the proof and a machine-checked Lean formalization on GitHub.

An independent group led by Russell Impagliazzo, Ramamohan Paturi and Ryan Williams will examine the proof. Hours later, Anthropic released its own refutation of ETH after its mathematician, Levent Alpöge, suggested OpenAI built on existing work. Williams said parts of the proof look routine, but two or three alien-looking parts need scrutiny.

Why it matters: If verified, the result would overturn a foundational assumption and force researchers to revisit years of work built on ETH.

• Source: Working with academic communities to build scientific understanding — OpenAI
• Source: Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean? — TechCrunch

3. Update: Nvidia and AMD Put $4 Billion Into Meta's AI Buildout

Nvidia will buy $3 billion of Meta's equity offering and AMD will buy $1 billion, Nvidia and AMD said Tuesday. Meta announced the $95 billion raise on Monday, as covered in our October 27 briefing. The offering combines $35 billion in underwritten sales with a $60 billion at-the-market program.

Nvidia said Meta will expand Vera Rubin and Vera CPU deployment and deepen work on Muse, Meta's AI agent. AMD said its Helios racks are running at Meta and its EPYC processors power Muse. Meta fell 1% on Tuesday, AMD rose 3%, and Nvidia was flat.

Why it matters: Chip suppliers are becoming investors in their own customers, a circular-financing pattern that critics say can obscure real demand.

• Source: Nvidia, AMD to invest $4 billion in Meta as tech giants deepen AI partnership — CNBC
• Source: Meta gets $3 billion from Nvidia and plans to spend more on Nvidia chips — The Decoder

4. UK Home Office May Face Probe Over Secret AI Use in Asylum

The Open Rights Group (ORG) published a report saying AI is embedded across the Home Office in near-total secrecy. Uses include detecting potential sham marriages, estimating asylum seekers' ages, and summarizing asylum interviews. The report also describes live facial recognition and AI-powered drones and satellites for border surveillance.

Wired reported that the Independent Chief Inspector of Borders and Immigration is considering a probe. The Home Office uses Asylum Case Summarisation and Asylum Policy Search, built on OpenAI's 2023 GPT-4. ORG found the summarization error rate was significantly higher than the published 9%, and the tools are absent from the public AI register.

Why it matters: Automated summaries can shape life-or-death decisions, yet asylum seekers are not told when the tools are used.

• Source: 'Digital hostile environment': AI secretly being embedded across Home Office, report claims — The Guardian
• Source: UK Home Office May Face Investigation Over Use of OpenAI Tech — Wired

5. Amazon Tells Engineers to Stop Chasing AI Token Use

Amazon published an internal memo Tuesday telling software engineers to stop treating token usage and code generation as measures of AI success. The memo, circulated in September, calls those metrics "deeply flawed" and says they can raise costs and complexity. Amazon shut down an internal leaderboard in May after employees used AI heavily to climb it.

A spokesperson said Amazon has not capped token usage. The memo says token and code metrics should be used "with care and judgment" alongside other data, and that long-term outcomes matter most. Marc Brooker, an Amazon Web Services (AWS) vice president, said the company looks at "feature velocity," including launches and successful production changes.

Why it matters: Counting AI usage rewards waste, not productivity, and Amazon is now telling engineers to use less when it does not help.

• Source: An update from Amazon engineering leadership on the role of the developer in the age of AI — About Amazon
• Source: Amazon's anti-tokenmaxxing memo says engineers should be pickier about using AI — CNBC

6. Apple Adds AI Assistants to Pages, Keynote and Numbers

Apple added AI assistants to Pages, Keynote and Numbers in beta. The assistants are built on Apple and OpenAI models, understand document context, and can draft text, design slides, and build tables and charts. The apps now support the Model Context Protocol (MCP) on macOS 27, letting external AI agents read, edit and act inside documents.

The update also brings Generate Fill to Pixelmator Pro, an AI assistant to Logic Pro, and MotionVFX content to Final Cut Pro. The features require an Apple Intelligence-capable device on iOS 27, iPadOS 27 or macOS 27, and are U.S. English only for now. Apple Creator Studio updates arrive in November as a free update for existing subscribers.

Why it matters: Apple is turning its productivity suite into an agent platform, not just adding a chatbot to each app.

• Source: Apple Creator Studio adds powerful new AI features and even more creative tools — Apple Newsroom
• Source: Apple is adding an AI chatbot to Pages, Keynote, and Numbers — The Verge

7. Disruptive Raises $10 Billion Fund for AI and Defense

Dallas-based Disruptive announced a $10 billion flagship fund Tuesday. It plans to back about 10 late-stage companies, with checks of up to $1 billion each. It is the firm's first blind-pool fund; it previously invested through deal-by-deal special purpose vehicles.

Past bets include Palantir, Airbnb and Groq. Bloomberg reported the firm told limited partners (LPs) it has returned more than $6 billion on roughly $1.9 billion invested. The Wall Street Journal first reported earlier this month that Disruptive was seeking the fund and had secured at least $7.5 billion in commitments.

Why it matters: The fund gives AI and defense startups another source of late-stage capital as public markets turn selective.

• Source: Disruptive Announces $10 billion Flagship Fund — PR Newswire
• Source: Disruptive raises $10B fund to make big bets on at least 10 companies — TechCrunch

8. AI Disinformation Floods Social Media Before US Midterms

AI-driven disinformation is spreading across social media ahead of the Nov. 3 midterms, according to experts quoted by The Guardian. Platforms have cut trust and safety teams and rely more on automated moderation. A false claim that Texas Senate candidate James Talarico was hospitalized in Ohio spread for days; his campaign released security video showing him at home.

The lie mutated in Spanish posts into a claim that he attempted suicide, the Digital Democracy Institute of the Americas (DDIA) found. In Pennsylvania, the Department of Homeland Security amplified a false claim about an immigrant working for Democrat Janelle Stelson's campaign. Meta requires disclosure for AI-generated election ads, YouTube has narrowed its rules, and X relies on Community Notes.

Why it matters: Corrections rarely catch every version of a lie, and platforms are doing less human review as the election approaches.

• Source: AI-driven disinformation running rampant on social media before US midterms, experts warn — The Guardian
• Source: Misinformation Against James Talarico Ramps Up in Texas — DDIA

9. Thieves Steal $100 Million in Data Center Gear Using GPS Spoofing

Thieves have stolen more than $100 million in data center freight since June, NBC News reported. The schemes start with hacked freight broker accounts and fake shipments. Trucking companies that respond receive malicious links that let thieves take over their accounts and bid on real loads.

Some thieves hire real drivers who do not know they are carrying stolen goods, or use AI-generated fake licenses. In two of the largest thefts, escorts were rammed and the trailers disappeared; one load carried $13 million in server equipment, another $10 million in rack components. GPS coordinates were spoofed in one case.

Why it matters: AI data centers depend on physical supply chains, and stolen servers and memory can end up on gray markets.

• Source: Highway thieves are spoofing GPS to steal millions in data center-bound cargo — Ars Technica
• Source: Highway robbers are targeting trucks carrying data center components — The Verge

10. Firefox Smart Window Can Fill Forms From Your Open Tabs

Mozilla added two features to Smart Window, its opt-in AI browsing mode in Firefox. The first can fill online forms using information from tabs the user selects. It appears when a form has at least four text fields, and Firefox does not submit the form. Data from selected pages is sent to Gemini 3.1 Flash Lite.

The second watches pages for changes such as a sale or new listing. Users type "/watch" and can track up to five pages per task, with up to 15 active tasks. Tasks run daily or weekly while the browser is open; monitoring uses the Qwen model. Both features are in beta on desktop in the U.S. and Canada.

Why it matters: Browser AI is moving from chat to tasks that touch personal data, making model choice and consent part of everyday browsing.

• Source: Skip the busywork with Smart Window — Mozilla
• Source: Firefox's Smart Window adds AI form filling and webpage monitoring — The Verge

AI Security: 5 Stories to Know Today

11. CyberStrikeAI Ran 44 AI Agents Against Japan Targets

TrendAI analyzed three attack infrastructures targeting Japan and found one run by CyberStrikeAI, an open-source AI penetration-testing platform. The platform split work across 44 agents, covering reconnaissance, exploitation and data theft. At least 19 organizations, including at least 16 in Japan, were probed in a short period in October.

Each target received a report in the same format, and the platform generated Japanese-language forms and fake login pages. It also throttled traffic to avoid detection. TrendAI found limited deep compromise; many attempts ended with "no session, no web shell, no exfiltration." The operator also targeted organizations in Taiwan and Hong Kong.

Why it matters: AI agents let one operator probe many targets at once, but the entry points were still exposed legacy systems and default credentials.

• Source: Not Only ARTEX: The Different Levels AI is Used in Attacks Targeting Japan — TrendAI
• Source: CyberStrikeAI Uses 44 AI Agents to Automate Hacking Operations Against Japan Targets — The Hacker News

12. Cloudflare Fixes Cross-Request Data Leak in GLM-5.3

Cloudflare disclosed that a configuration error and an SGLang bug exposed data across requests on its GLM-5.3 Workers AI model. Between Aug. 27 and Sept. 9, 155 responses contained data from other customers' requests. All involved prompts longer than 681,280 tokens. The model had a 1,048,576-token context window, but the prefill key-value (KV) cache pool held only 681,280 tokens.

A rejected over-length request still moved through the disaggregated prefill/decode handoff, so decode read stale cache pages. Cloudflare took the model offline and enlarged the pool. It patched the engine and restored service the same day. It found no evidence of deliberate exploitation. SGLang released fixes in v0.5.20 through v0.5.22 and recommends operators upgrade.

Why it matters: Multi-tenant AI services can leak one customer's prompt into another's answer without any error, so cache isolation needs continuous testing.

• Source: Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang — Cloudflare
• Source: SGLang security update: improving request isolation — SGLang

13. Warp Breach Exposes AI Agent Logs and Telemetry Data

Warp disclosed that attackers breached a third-party vendor in its infrastructure and accessed agent traces and telemetry data. CEO Zach Lloyd said the company detected the operation within hours and estimates an upper bound of roughly 1% of Warp users may have been affected. Warp says it has at least 800,000 users.

Lloyd described the attack as an "agent swarm" using a sophisticated penetration and escalation technique, likely starting with a vendor dependency. The company engaged CrowdStrike, notified affected users and contacted law enforcement. Warp did not say whether source code, prompts or credentials were exposed.

Why it matters: AI coding tools hold source code, prompts and credentials, so a vendor breach can expose far more than chat logs.

• Source: Warp breach statement — Zach Lloyd on X
• Source: Warp discloses breach exposing AI agent logs, user telemetry data — BleepingComputer

14. OpenAI Fixes Flaw That Leaked Hidden Reasoning Traces

Xint researchers showed that OpenAI's compaction feature could copy a model's hidden reasoning as plain text. OpenAI's reasoning models return raw reasoning only as an encrypted item; compaction shrinks a conversation into another encrypted item. Xint asked the compaction step to preserve a prior reasoning item word for word, replacing spaces with tabs to evade filters.

It then sent the compaction to cheaper models, which repeated the text. OpenAI confirmed the issue, fixed it on Oct. 23 and paid a $6,000 bounty. Xint tested gpt-5.6-sol and did not check GPT-6 Astra. It could not prove the copied text was word-for-word the original reasoning.

Why it matters: Reasoning traces are valuable training data, and any server-side step that reads them can become a decryption oracle.

• Source: Stealing OpenAI Reasoning Traces via Compaction Smuggling — Xint
• Source: OpenAI Compaction Flaw Let Researchers Bypass Hidden Reasoning Protections — The Hacker News

15. Aokigahara Stealer Targets AI Assistants and Developer Keys

Check Point researchers found a new infostealer, Aokigahara, that targets AI assistants and developer tools alongside browsers and crypto wallets. It collects credentials, session data and configuration files from Claude, ChatGPT Desktop and Cursor, among other AI tools. It also grabs Git credentials, Secure Shell (SSH) private keys and cloud configuration files.

Stolen AI sessions can be resold or used to access paid services, private conversations and source code, Check Point said. The malware checks for virtual machines, sandboxes and debuggers, then encrypts stolen data before sending it to its command server. Check Point said it has not confirmed how operators monetize the data.

Why it matters: An AI assistant is now a keyring, and one infected laptop can expose code repositories, cloud accounts and chat histories at once.

• Source: Infostealer riding the AI Train — Check Point Research
• Source: 'Aokigahara' Stealer Targets 'Developer Keyrings' & AI Assistants — Dark Reading

More AI Stories Worth Reading

• China's STAR 50 technology index has fallen 40% since June, wiping $2.4 trillion from Chinese stocks as AI profits lag. — The Guardian
• Opera introduced WebAI, an early-preview standard that lets open-weight models run locally in its browser instead of the cloud. — TechCrunch
• Nvidia backed Indian startup Simplismart in a $14 million round for software that helps companies run AI models with fewer GPUs. — TechCrunch
• Publishers including Axios and Every are building feeds and articles for AI agents, not just human readers. — Wired

Related Reading on Kill The AI

• Anthropic filed for a Nasdaq IPO, Meta announced a $95 billion stock sale, and Microsoft offered buyouts to 60,000 employees.
• ARC Prize rejected OpenAI's 99.9% GPT-6 Astra score, Elvis rights holders sued Anthropic, and OpenAI offered 90% discounts.
• OpenAI cut training from the live internet, Kimi K3.1 Pro neared Opus 5.5, and the UK pledged £100M for biosecurity.
• Our data review of the AI boom's first eight weeks, from capex and chips to layoffs.

Methodology & Sources

Compiled October 28, 2026, using multi-source research. Primary sources were Google DeepMind, OpenAI, About Amazon, Apple Newsroom, PR Newswire, DDIA, Mozilla, TrendAI, Cloudflare, SGLang, X, Xint and Check Point Research. News outlets were Ars Technica, TechCrunch, CNBC, The Decoder, The Guardian, Wired, The Hacker News, BleepingComputer, Dark Reading and The Verge. All sources are free to read; no paywalled links. Details are as reported at compilation time.

Frequently asked questions