Friday, and the AI news this week came with receipts. A federal judge in Mississippi told xAI and the Justice Department that the NAACP can keep suing over xAI's gas turbines, OpenAI shipped a hacking model and paired it with some numbers worth chewing on, and Uber admitted that almost nobody is booking rides with AI agents. Let's get into it.
01 of 03
xAI loses its bid to kill the NAACP turbine lawsuit
On October 30, U.S. District Judge Debra M. Brown of the Northern District of Mississippi denied xAI's motion to dismiss a Clean Air Act lawsuit brought by the NAACP and its Mississippi State Conference. Earthjustice and the Southern Environmental Law Center are representing them. The defendants are xAI, which is now part of SpaceX, and its subsidiary MZX Tech.
The fight is over the Colossus Gas Plant in Southaven, Mississippi, which powers xAI's Colossus 2 data center just across the state line near Memphis. The suit says xAI put in 27 gas turbines, with at least 495 megawatts of combined capacity, between August and December 2025 without permits. That breaks down to 14 SMT-130 units, one M35, and 12 GE TM 2500 units. The complaint estimates potential emissions of about 1,700 tons of nitrogen oxides a year, plus 580 tons of carbon monoxide, 180 tons of particulate matter, and 19 tons of formaldehyde. The NAACP wants an injunction, pollution controls, and civil penalties of up to $124,426 per day per violation.
xAI threw four arguments at the wall. The turbines sit on flatbed trailers, so they are "mobile sources" and exempt. Mississippi's environmental agency sent a July 2025 letter treating them as exempt. It did not get fair notice. And citizen lawsuits under the Clean Air Act are unconstitutional because they step on the president's enforcement power. The Justice Department jumped in back in June on xAI's side, arguing the case could cut power to an AI facility important to the economy and the "Department of War."
Judge Brown did not buy any of it. A turbine that is not self-propelled counts as stationary, trailer or not. The state's letter does not bind a federal court. And citizen suits stand, because the plaintiffs "sue on their own behalf, in their own name." The ruling does not say xAI broke the law. It just means the case moves forward, and the NAACP's request for a preliminary injunction is still pending.
Here is where I land. Putting a turbine on a trailer to dodge a permit is the kind of trick that works right up until somebody hauls you into court. Neil Osnato of Persistence Analytics Group put it better than I can to TechTarget: "A trailer can make a turbine portable. It does not necessarily make its regulatory obligations temporary." The bigger deal is the DOJ angle. The federal government tried to argue regular folks should not be allowed to enforce clean air law against an AI company. A judge said no. That matters way past Southaven.
Sources: TechTarget, The Next Web, MLex
02 of 03
OpenAI ships GPT-6 Cyber, and the price tag tells the story
OpenAI announced GPT-6 Cyber on Thursday. It is a red teaming model built to find software vulnerabilities, write exploits to prove them, and help patch them. OpenAI says it is a big step up from GPT-5.6 Cyber and beats its general GPT-6 Astra model on most cyber tasks, even when Astra does not refuse the request.
You cannot just sign up for it. It goes first to a limited group in Daybreak Red, the top tier of OpenAI's cybersecurity program, meant for offensive security researchers. The lower tier, Daybreak Blue, gives organizations models for finding and fixing bugs. OpenAI says "thousands" of security pros are in Daybreak, and it announced a $1 billion initiative to support members working on critical infrastructure like the power grid and water systems, plus a partnership with industrial security firm Dragos.
The numbers OpenAI is touting: TechCrunch reports OpenAI said the model was used to scan the Linux kernel and prepare patches, finding 948 potential issues. CyberScoop adds that GPT-6 Cyber found at least 81 new Firefox vulnerabilities and a critical remote code execution bug in Brotli, the open-source compression library. All of it was reported and patched. On guardrails, OpenAI says there are tighter identity checks, more frequent check-ins with OpenAI when the model is asked to do high-risk stuff, and extra monitoring for rule-breaking. Researcher Eric Wallace said the model is trained to be good at "listening to your instructions on what you should touch and what you should not touch."
Not everybody is clapping. Max Justicz, CTO of Charge Robotics, told TechCrunch he found a serious Linux bug using a GPT-5.6 model through Daybreak Blue, that he joined by uploading a photo of his driver's license, and that the Blue tier got "nerfed." His warning: "We should expect some extremely crazy bugs to be found and exploited over the next few months." Daybreak Blue users are also mad that OpenAI told them session data from personal accounts will be used "to improve our models," even if they opted out. OpenAI safety lead Fouad Matin says that data is for abuse prevention.
What I keep coming back to is one line in the TechCrunch piece. The Linux bug hunt cost OpenAI about $1.2 million. It spent about $6 million in tokens on a fluid dynamics math problem. If you believe your own models are about to make hacking cheap for everybody, and OpenAI clearly does, then spending five times more on a math flex than on hardening the software the whole internet runs on is backwards. Good model. Wrong priorities.
Sources: OpenAI, CyberScoop, TechCrunch
03 of 03
Four rides: the AI agent reality check
On Uber's third-quarter earnings call Wednesday, CFO Balaji Krishnamurthy said Uber handled around 320 million trips and orders across Uber and Uber Eats during the week of Halloween, its busiest week. AI agents placed 500 of the Uber Eats orders. They booked four rides. Four.
CEO Dara Khosrowshahi did not sugarcoat it: "It is very, very early and very limited traction that we are seeing so far." He also pointed out that booking a ride in the app takes about 50 seconds on average and reordering a favorite meal takes about 90, which does not leave an agent much time to save you. Uber is not alone. DoorDash said on its own earnings call last week that it has gotten only "a few hundred orders" from AI agents.
Same day, Bloomberg reported that Apple is cutting what its own people can spend on AI. In October it dropped the daily AI spending cap for engineering teams from $300 to $200, with some allowances reportedly headed to $100 a day. Outside engineering, Apple is reportedly weighing a cap as low as $25 a day. Apple had not commented to TechCrunch. TechCrunch framed it as the hangover from this year's "tokenmaxxing" craze, when companies pushed staff to burn as many tokens as possible, then the bills showed up.
I am not here to dunk on agents. But these two stories together say something. The agent pitch was that everybody would hand off errands to a bot. Turns out people can already order a burrito in 90 seconds and do not need help. Where agents pay off is the boring, expensive work inside a business, and even there Apple is checking the receipt. That is healthy. Spend on AI where it clearly earns its keep, and stop pretending a chatbot calling an Uber is the future.
Sources: TechCrunch on Uber, Uber Q3 call transcript, Bloomberg on Apple, TechCrunch on Apple