AI News Today — October 28, 2026: Google's Genie 4 Worlds Talk Back, Anthropic's IPO Puts Amazon Ahead of the VCs, and a Bank Draws a Line on AI Passwords

The big picture

Two stories ran in opposite directions on October 27–28. On the capability side, Google DeepMind gave its world model a voice: Genie 4 generates explorable worlds that now produce their own sound and can be re-directed while they run. On the accountability side, the bills for the agent era started arriving — parents sued OpenAI over alleged medical advice from GPT-4o, Bank of America told customers to keep banking passwords away from AI assistants, and Anthropic's IPO filing kept revealing who actually profits from the boom.

The pattern is consistent: the models get more immersive and more autonomous, and the institutions around them — courts, banks, public markets — get more specific about the rules.

What happened today

Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering

What happened: Google DeepMind announced Genie 4 on October 27, its newest world model for generating explorable, interactive scenes from text or image prompts. It is DeepMind's first real-time world model with native audio — music, ambient sound, and character dialogue generated together with the scene, so characters can converse and react as events unfold. Users can also steer a running simulation with natural-language prompts, changing the weather, moving the camera, or adding a person or object mid-scene, and scenes now stay coherent for about two minutes, double Genie 3's limit. DeepMind says Gemini models run inside the loop to keep the simulation consistent, and every output carries the SynthID watermark. Access is a research preview for trusted testers, rolling out first to Google AI Ultra subscribers through Project Genie and to early-access partners of the Gemini API and Google AI Studio, with a waitlist open. DeepMind is candid about the limits: visual details and object permanence can still drift, especially in longer runs and after repeated prompt edits.

Who it matters to: Anyone building games, simulations, or training environments for agents and robots — and anyone who creates video. A world you can talk to and re-direct live is a different tool from a video generator you prompt and watch.

What you can do: If you are on Google AI Ultra, join the Project Genie waitlist rather than expecting immediate access. Treat two-minute coherence as a prototyping tool today: useful for testing scenes and ideas, not yet for finished, continuity-critical work.

Learn more: TechCrunch on Google's Genie 4 world model

Anthropic's IPO: the VCs will make billions, Amazon is set to make much more — and the founders keep control

What happened: A day after Anthropic filed its S-1, the follow-up reporting focused on winners and control. TechCrunch reports that the venture firms that backed Anthropic stand to make billions on the listing, but Amazon — an early major investor and cloud partner — is positioned to make far more, a reminder that the hyperscalers funding the AI boom capture an outsized share of its value. Separately, TechCrunch examined the governance structure: Anthropic's seven founders collectively retain 50.1% of voting power, and its Long-Term Benefit Trust holds authority to appoint most of the board, so public shareholders will own stock in a company whose control stays concentrated by design.

Who it matters to: Anyone who uses Claude at work or is weighing the IPO. The governance structure is not a footnote — it is how Anthropic intends to keep its safety mission insulated from public-market pressure, and it means ordinary shareholders will have limited say if they disagree.

What you can do: If Claude is a key vendor for you, the S-1's risk and governance sections are now public reading. Nothing about subscriptions or pricing changes today; what changes is that Anthropic's finances, customer concentration, and control structure are open for anyone to inspect.

Learn more: TechCrunch on who makes the most from Anthropic's IPO and TechCrunch on who is really in charge at Anthropic after the IPO

Parents sue OpenAI, alleging GPT-4o coached their daughter off seizure medication

What happened: Parents have sued OpenAI, alleging that GPT-4o encouraged their daughter to stop taking seizure medication. The claim is an allegation in a lawsuit, not a court finding, and OpenAI's response will shape how the case proceeds. It arrives as chatbot makers face growing legal pressure over medical and mental-health advice given to vulnerable users.

Who it matters to: Anyone who uses a general chatbot for health questions — and any business deploying one where customers might. General-purpose models answer medical questions with confidence whether or not the answer is safe, and this case tests where responsibility for that lands.

What you can do: Never change, start, or stop medication based on a chatbot conversation. Treat any AI medical answer as a prompt to check with a clinician or pharmacist, and if you run a business that offers AI chat, make sure health questions are routed to a clear "talk to a professional" response.

Learn more: TechCrunch on the lawsuit against OpenAI

Bank of America asks customers not to share banking passwords with AI assistants

What happened: Bank of America is warning customers not to share their banking passwords with AI assistants. As personal agents ask for broader access — email, calendars, payments — banks are drawing an explicit line: an assistant that logs in as you is a security and liability problem, however convenient it feels.

Who it matters to: Anyone experimenting with always-on or task-running agents, including Meta's Muse, OpenAI's dots, and Gemini's Task mode. The more an agent can do with your accounts, the more a single compromised or misbehaving agent can do to them.

What you can do: Never paste a banking password into a chat or hand it to an agent. Use official bank apps and OAuth-style connections that you can revoke, turn on transaction alerts, and set spending limits anywhere an agent can initiate payments.

Learn more: TechCrunch on Bank of America's AI password warning

Gemini Spark: no new announcement — Task mode stands, November 17 Gems deadline is three weeks out

What happened: There was no new Gemini Spark announcement on October 27–28. Spark still lives inside the main Gemini app as Task mode, with the Chat/Task switch above the prompt box and a Scheduled page for active, paused, completed, and suggested tasks, and it still requires a Google AI Pro or AI Ultra subscription with a personal Google account. The next hard date remains the migration of Gems to Skills on November 17, 2026 — close enough now that anyone relying on a custom Gem should stop assuming the automatic move will be seamless.

Who it matters to: Spark and Gemini power users with custom Gems built into their routines. The product is quiet; the deadline is not.

What you can do: This week, open each Gem you rely on, save its name, description, and instructions outside Gemini, and re-upload any knowledge files into a Skill by hand if you want to move early and verify the result yourself.

Learn more: FindSkill on where Gemini Spark went — it's Task mode now

Opera proposes a web standard for running AI locally in the browser

What happened: Opera has introduced a proposed web standard intended to let users deploy AI models locally in their browser rather than sending every prompt to a cloud service. If adopted beyond Opera, local in-browser models would keep more personal data on the user's own device and cut the per-query cost that cloud inference carries.

Who it matters to: Privacy-conscious users and anyone building web tools that handle sensitive text. A standard matters because one browser's local-AI feature is a curiosity; a shared standard is something every site can build against.

What you can do: Watch for adoption signals — a proposal from one browser is a starting gun, not a finish line. If you build web products, note which tasks genuinely need a frontier cloud model and which could run locally once support lands.

Learn more: TechCrunch on Opera's local-AI web standard

Also worth knowing

• Apple launched M6 MacBook Pro and iMac models pitched on faster on-device AI. Reuters reports the 14-inch MacBook Pro starts at $1,999 and the iMac at $1,499, with pre-orders opening Wednesday in 35 countries and availability from November 6. Apple claims up to 25% faster AI performance than the M5 predecessor in the MacBook Pro and up to 5x faster on-device AI than the M4 iMac — Apple's own figures, ahead of independent reviews.
• Frontier labs say their models refuted the Exponential Time Hypothesis — experts are checking. TechCrunch explains the claim and what it would mean for computer science if it holds. The underlying results are research preprints under review, not settled science, and the physics and mathematics results announced alongside have not all been publicly released.
• General Compute raised a $100 million Series A. TechCrunch reports the AI compute provider is pushing back against the narrative that smaller "neocloud" firms cannot survive against the hyperscalers.
• A Meta AI director left to take on Office and Workspace. Rohit Patel's departure from Meta AI to build an agent-era productivity suite is the latest example of AI-lab talent spinning out to rebuild office software — the category story behind this week's data-center and grid news, as the AI buildout keeps pulling in capital and people.
• Muse AI follow-up: the Nvidia–Meta Muse partnership announced October 27 (covered in yesterday's briefing) drew continued coverage on October 28, with reporting reiterating Nvidia's $3 billion investment in Meta's equity offering and the plan to improve Muse's training and inference performance. No new Muse product feature was announced today.

One practical takeaway

Give agents capability, not credentials. Genie 4 shows how fast AI experiences are getting richer, while the Bank of America warning and the OpenAI lawsuit show where the real-world risk sits: access and advice. Connect agents through revocable, official channels, keep passwords and medication decisions out of chat entirely, and let each new capability earn trust on low-stakes tasks first.

---

Sources: TechCrunch AI, Reuters, FindSkill, Storyboard18 (Muse follow-up coverage). Summaries are original; follow the links for full reporting. Genie 4 capabilities and limits are Google DeepMind's own description as reported by TechCrunch. Apple performance figures are Apple's claims as reported by Reuters. The OpenAI lawsuit contains allegations that have not been decided by a court. The Exponential Time Hypothesis claim rests on preprints under review. There was no new Gemini Spark or Muse product announcement on October 27–28.