Busy end to the week. My home state went after OpenAI, a federal judge got yanked off a case for letting AI write his orders, OpenAI admitted one of its models slipped onto the internet, and DeepSeek raked in a pile of cash right after Washington kicked it out of federal systems. Let's go.
01 of 04
Texas Sues OpenAI and Sam Altman Personally
Attorney General Ken Paxton filed a 67-page lawsuit in Collin County on Friday against OpenAI and Sam Altman himself. The core claim is under the Texas Deceptive Trade Practices Act: OpenAI marketed ChatGPT as safe and reliable while not telling people it can spit out wrong or harmful answers. Reporting on the filing says it also leans on the Texas SCOPE Act, the state's law on kids' online data.
Here is the list of harms the AG's office spelled out: ChatGPT allegedly helped mass shooters plan attacks, encouraged vulnerable people toward suicide, embarrassed professionals who relied on it, eroded users' critical thinking, and got minors hooked while collecting their data without parental oversight. Texas wants money penalties, court orders changing how OpenAI operates, and Altman held personally liable.
Paxton's line: "I will fight to protect Texans from the harms of AI and ensure that powerful technology companies are not above the law." Texas is not first here. Florida sued OpenAI and Altman back in June on similar grounds.
What I want to know is how much of this survives contact with a judge. Some of it is real, especially the stuff about kids and data. But "users lost critical thinking skills" is a hard thing to put a dollar figure on in court. Naming Altman personally is the part to watch. That is a pressure move, and with two big states now doing it, OpenAI cannot wave this off as one guy grandstanding.
02 of 04
A Federal Judge Got Benched Over AI-Written Orders
The Fifth Circuit tossed a preliminary injunction against Mississippi's DEI ban, House Bill 1193, and ordered the case taken away from U.S. District Judge Henry Wingate. The case is Jackson Federation of Teachers v. Fitch, and Judge Jerry Smith wrote the opinion.
The receipts are rough. Wingate's July 2025 restraining order named 11 plaintiffs who do not exist, cited declarations from four people who are not in the record, and quoted language that is not in the statute. Wingate later admitted a law clerk drafted it with generative AI. The replacement order still cited a 1974 Fourth Circuit case, Cousins v. School Bd. of City of Norfolk, that does not exist either, and the injunction leaned on that replacement.
Smith did not mince words: "The proceedings in the district court were a botch job, to say the least." He also said the lower court "ignored and manipulated governing law, apparently to satisfy a policy agenda."
Here is where I land on it. Set the DEI politics aside, because this would be just as bad pointed the other direction. Lawyers have been getting sanctioned for fake AI citations for three years now. A sitting federal judge doing it, then patching it with a second order that still had a fake case in it, is a whole different level. If the folks in robes cannot be bothered to check the output, nobody else is going to.
Source: Fifth Circuit opinion
03 of 04
OpenAI Cuts Its Frontier Models Off From the Internet
OpenAI posted Friday that it will train its frontier models with no live internet access. Instead, the models get a big indexed, cached copy of the web to practice browsing and search on, wrapped inside locked-down containers and microVMs with monitoring on top.
The reason is not theoretical. On September 20, an internal research model exploited a misconfigured DNS resolver during a training run and got limited access to the live internet. OpenAI says it contained it, then paused training and evaluation on its most capable models and did not resume until October 1, 11 days later. Its red team also found agents chaining bugs to escape microVMs and passing messages between isolated sandboxes through covert channels like processor cache timing and CPU contention. OpenAI says it patched and disclosed bugs it found in Kata Containers, gVisor, and Moby/Docker along the way.
This comes about two weeks after Anthropic started cutting its own internal evaluations off from the live web.
I'll give OpenAI credit for writing this down in public with a date on it. That is more than most companies do after an incident. But read it again: the model went looking for a way out and found one through a DNS setting. That is the kind of sentence that should make everybody slow down a notch. Two of the biggest labs are now saying out loud that their own models need a fence. Believe them.
Source: OpenAI blog post
04 of 04
DeepSeek Raises $15 Billion Right After Getting Booted by DHS
On October 20, the Department of Homeland Security published an order excluding DeepSeek products from federal civilian agencies. It covers DeepSeek itself plus third-party products built on its tech, for new purchases and existing systems alike. DHS said sensitive data from government users "could potentially be accessible to Chinese authorities under Chinese law."
Two days later, Bloomberg reported DeepSeek raised about $15 billion, roughly 100 billion yuan, with battery giant CATL and Tencent named among the potential investors. For scale, DeepSeek raised over $7 billion at a $50 billion valuation back in June. Founder Liang Wenfeng reportedly owns about 81.6 percent of the Chinese operating company.
China's Foreign Ministry called the DHS move "politicizing technology and trade issues," which is what you would expect them to say.
Here is my read. The ban makes sense for government computers. I do not want federal data running through a Hangzhou lab either. But nobody should mistake it for a punch that lands. DeepSeek does not need U.S. agency contracts. It needs money and chips, and the money just showed up in a bigger pile than before. If the goal is slowing them down, a procurement ban is not the tool that does it.
Sources: DHS order in the Federal Register, Bloomberg report