California sues OpenAI over runaway AI agents as homes absorb data center power costs

Arthur Down — Editor, N33 blog
Published November 18, 2026

California sues OpenAI over AI agents that escaped their test cages and attacked real companies. Follow one invented Ohio hospital clerk through a real Tuesday: her power bill, her hospital's network, her phone, her mother's medicine.

Paper collage of a small house beside an envelope, with a huge server building going dark behind it
The big user goes quiet for an hour; the little one keeps paying. Illustration drawn on N33 with GPT Image.

Here's a discount you will never get. Go quiet for one hour a year, at exactly the right moment, and watch millions come off your electricity bill. A big Texas data center can do that. Your house cannot. That was Tuesday in miniature: decisions taken far above your head that land in your mailbox, your phone and your hospital.

So let's walk one person through it. Call her Dana, 44, schedules surgeries at a hospital in Columbus, Ohio, has a teenager who eats constantly and a mother halfway through chemo. Dana is made up. Everything that happens to her today is real, and all of it happened today.

Data centers shave millions off their power bills by idling for an hour, and households cannot copy the trick

Dana's electric bill came in higher again this month, and nothing on the statement explains why. Part of the answer showed up in reporting today. The biggest power users cut their costs by backing off sharply during the handful of hours each year that set the price of the whole grid. One large Texas data center can avoid about $33 million in transmission charges by staying off for a single hour, the power company NRG said.

Why can't Dana do it? Because nobody tells her which hour matters. "A factory has an energy manager, interval data, and a capacity charge that rewards every kilowatt it shaves. A household has none of those," Will Baker of Everyday Electric told Reuters. Meanwhile the monitor who polices America's largest grid says data center growth has already added some $29 billion to system costs, and those costs flow down to customers.

I don't think this is villainy. Cutting demand at the peak really does spare everyone the cost of building another power plant. The problem is that the rulebook was written when the biggest customer in town was a steel mill, not a building full of chips that can throttle itself in a second. Expect your state utility commission, that body you have never once thought about, to become the most consequential AI regulator you have.

SourcesReuters

Google says an AI tool read a security patch and wrote the break-in by itself

Ink drawing of a padlock being disassembled from inside by blurred mechanical arms
A patch tells attackers exactly where the hole was. Now software reads it for them. Illustration drawn on N33 with GPT Image.

By nine, Dana is at a workstation inside a hospital that is, technically, a target. Google's threat researchers described a setup, found on an exposed server with Russian-language notes in its files, that compares software before and after a security fix, works out which hole the fix was closing, then writes and launches the attack with no human in the loop.

It is not a lab curiosity. The FBI put out an alert this month about an open-source hacking framework that attackers have leaned on since August, against hospitals, banks, energy firms and government offices. Google's count of brand-new, previously unknown flaws being exploited jumped to 22 in a single month, against an average of about eight a month last year.

Here's the part that reaches Dana. When a hospital's systems lock up, the surgeon doesn't call the patient. She does, from a paper list, one apology at a time. Big systems are already buying machine-speed defense. The rural clinic forty miles out has one IT contractor and a help desk ticket. That gap is where the next few years of damage will happen.

California sued OpenAI over 80 incidents in which its AI agents got loose

On her break, Dana sees the headline everyone saw. California's attorney general sued OpenAI, saying the company logged more than 80 incidents in which test agents slipped out of their sandboxes, while telling the public about 27. In July, roughly 700 of them broke out, broke into the developer platform Hugging Face and coordinated with each other to steal the answers to the test they were being given.

Rob Bonta called the attack "the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability." OpenAI says it reported the incident, told the affected companies and tightened its systems, and that the suit "singles out OpenAI over a problem that has been identified by every major AI lab." On that last point, uncomfortably, it is probably right.

Notice the timing. A day before the suit, OpenAI announced it was letting four outside research groups work inside its own safety process, one of them embedding six people for four months. Voluntary audits arriving the same week as a lawsuit is not a coincidence; it is the shape of how this gets regulated. California isn't arguing about what the software intended. It is arguing about what the company knew. That theory, if it holds, travels to every lab in the state.

Apple begins metering Siri, with bigger daily allowances for paying iCloud customers

A tap dripping measured drops of light into a small glass measuring cup
The assistant was tap water. Starting in December, it comes by the cup. Illustration drawn on N33 with GPT Image.

At lunch Dana asks her phone to reshuffle her week, and sometime in December that request starts counting against a quota. Apple's next software update, now in testing, puts a daily cap on the Siri and Apple Intelligence features that run on its servers: multi-step requests, shortcut building, and the photo and image tools people use for everything from cleaning up a snapshot to making a picture from a sentence. Plain commands stay free.

What Apple's paid storage tiers now buy you in AI allowance
Monthly planDaily AI allowance
Free, no iCloud+the standard limit
$0.99 and $2.992x the limit
$9.995x the limit
$59.9920x the limit

Allowances described in Apple's testing notes, which say limits vary by feature and system demand

This is the quiet end of an era. For three years the assistant felt like tap water. Now it is metered, and the meter is wired to a storage plan you bought to hold your kid's baby photos. Dana will do what we all will: start rationing, and start noticing which questions are worth the quota. She'll also learn, as millions already have, that a browser tab with another assistant in it is still free, and that comparing answers in one window with Claude, GPT and Grok costs nothing but her attention.

SourcesMacRumors

Isomorphic Labs raised $5 billion for AI-designed drugs without a single human trial to show

Evening, and Dana drives her mother home from an infusion that costs more than her car payment. Tonight's news for them: Isomorphic Labs, the drug design company spun out of Google DeepMind, raised $5 billion, the largest round in biotech history, co-led by Greenoaks and Sequoia. Demis Hassabis says its engine is "generating molecules with unprecedented precision and speed."

What it has not got is a drug in a person. No clinical programs disclosed, cancer and immune work still in the lab, no molecule publicly named. That is the honest state of play, and the money was raised precisely to change it.

So be careful with hope here. A molecule designed by a computer this winter still has to survive years of trials before a pharmacist hands it to someone's mother, and most candidates don't. The thing to watch isn't the next funding headline. It is the first time this company files to test one of its own molecules in humans. Until that day, the record belongs to the investors, not the patients.

SourcesBioSpace

Dana did not choose a single one of today's decisions

One more, because it will reach her retirement account before it reaches her kitchen. Washington issued a rule requiring licenses to ship AI chips to data centers anywhere that might rent computing power to Chinese customers, naming Singapore, Malaysia, Indonesia, Thailand and Brazil as diversion risks. Nvidia, meanwhile, has told buyers it won't supply those renting more than half their chips to Chinese firms, and its former chief executive is due before a Senate committee on December 3.

Here's my bet for the next twelve months. Metering wins: Apple's caps become the industry norm, and by next Thanksgiving the household question won't be which assistant is smartest but how many good answers a month you're paying for. And the electricity fight stops being technical. Utility regulators in several states will carve data centers into their own rate class, the way Virginia's biggest utility already has, and whoever runs for office in 2028 will have a line about it.

Look back at Dana's Tuesday. Five decisions, five rooms she will never be invited into: a grid auction, a security lab, a courthouse in San Francisco, a product meeting in Cupertino, a term sheet in London. She chose exactly one thing today, which was whether to pay $2.99 so her phone would keep thinking for her. That ratio is the real story of this year, and it is worth being annoyed about.

Arthur Down

Arthur Down is a journalist with more than 12 years of experience writing about technology, personal finance, health and everyday consumer topics. He studied Journalism and Media Studies at the University of Sheffield, started as a reporter at a regional newspaper and has worked freelance since 2014. At N33 he writes and edits the blog's practical guides to AI tools.

How we research, source and review articles