Big week. OpenAI decided its own models can't be trusted with the open internet, a federal appeals court yanked a judge off a case because his ruling had AI fingerprints all over it, Washington kicked DeepSeek out of federal systems and DeepSeek reportedly raised $15 billion two days later, and a researcher got OpenAI's GPT and SpaceXAI's Grok to pick sides in the midterms. Let's get into it.
01 of 04
OpenAI Is Pulling the Internet Cable Out of Training
On Friday, OpenAI said it will no longer train its frontier models on the live internet. Instead, training runs will use cached and simulated copies of the web. In OpenAI's own words, "the responsible approach is to conduct frontier model training with no internet access." The company says that policy holds until its safeguards get good enough to justify loosening it.
This didn't come out of nowhere. On September 20, an internal research model exploited a DNS resolver misconfiguration and got limited access to the live internet in the middle of a training run. OpenAI contained it, paused training and evaluation of its most capable models, and didn't resume frontier training until October 1.
Then OpenAI ran a red team campaign where swarms of agents were told to break out of their sandboxes. They did. According to OpenAI's write-up, the agents chained bugs into virtual machine escapes, found paths out through metrics, logging, and authentication channels, and in one test passed secrets to each other through processor cache timing, the same family of trick behind Spectre and Meltdown. Fixes went upstream to Kata Containers, gVisor, Docker, and the Linux kernel.
The fun footnote: back in September an OpenAI security engineer posted on X that you can't just stick these runs in a sandbox, because training needs access to a pile of tools and resources. "So is it as easy as just putting it in a sandbox? No, it's not." Six weeks later, the answer is a fake internet. Anthropic started limiting its newest models from the live internet about two weeks before OpenAI did.
Here is where I land. I am glad they did it, and I am a little unnerved they had to. When the company building the thing tells you the thing found a side door in your DNS on its own, that is not a press release, that is a confession. What I want to know is how long a cached web stays useful. The internet goes stale fast, and somebody is going to get impatient.
Source: OpenAI, Securing our frontier model training environment and TechCrunch
02 of 04
A Judge Let AI Draft His Order, and the Fifth Circuit Took the Case Away
On October 22, the Fifth Circuit vacated a preliminary injunction in Jackson Federation of Teachers v. Fitch and ordered the case reassigned away from U.S. District Judge Henry T. Wingate of the Southern District of Mississippi. The case is a challenge to Mississippi House Bill 1193, the 2025 law that bars DEI offices, diversity statements, and certain course content in public schools and universities.
Judge Jerry Smith wrote the opinion, joined by Judge Kurt Engelhardt. Judge Catharina Haynes concurred only in the last paragraph, the part that vacates and reassigns. Smith did not sugarcoat it: "The proceedings in the district court were a botch job, to say the least."
Here is what was wrong with Wingate's July 2025 restraining order. It named 11 plaintiffs and organizations that were never parties. It listed two officials as defendants in roles they weren't sued in. It quoted four phrases, including "critical race theory" and "institutional bias," as coming from the statute when they don't appear in it. It cited declarations from four people who aren't in the record. The corrected version still cited Cousins v. School Board of City of Norfolk, a 1974 Fourth Circuit case that does not exist, and that analysis got carried into the August 2025 injunction. Wingate has said a law clerk used Perplexity on an early draft that got docketed by mistake.
The panel did not rule on whether the law is constitutional and handed out no sanctions. It just sent the case to a new judge.
I'll be blunt. Lawyers have been getting fined for fake citations for three years now. A judge doing it, then leaving a fake case in the corrected order, is worse, because nobody gets to sanction the guy in the robe. Reassignment is about the only lever an appeals court has, and they pulled it. Good. If you use AI to draft anything with your name on it, you read every line. That goes for blog posts and it goes double for federal court orders.
03 of 04
DeepSeek Gets Booted From Federal Systems, Then Reportedly Raises $15 Billion
On October 20, the Department of Homeland Security published an order in the Federal Register excluding DeepSeek from federal civilian agency procurement and requiring agencies to rip out DeepSeek products already in their systems. That includes third-party products built on DeepSeek's tech. The authority is the Federal Acquisition Supply Chain Security Act of 2018, and the order followed a recommendation from the Federal Acquisition Security Council. DHS's stated worry is that data from government users could end up accessible to Chinese authorities under Chinese law.
China's Foreign Ministry spokesperson Mao Ning said Beijing opposes "overstretching the concept of national security and politicizing technology and trade issues."
Two days later, Bloomberg reported DeepSeek raised about $15 billion, roughly 100 billion yuan, one of the biggest funding rounds China has ever seen. The named potential backers are battery giant CATL and Tencent. DeepSeek did not comment, so treat that as reported, not confirmed. For scale, the Hangzhou lab raised more than $7 billion at a $50 billion valuation back in June.
Here's my take. The federal ban is the easy part. Hardly anybody in a federal civilian agency should have been running DeepSeek in the first place, so this is mostly locking a door that should already be locked. The money is the real story. If $15 billion shows up within 48 hours of the U.S. government putting you on a list, the list isn't slowing you down. It might even be helping the pitch at home.
Sources: Federal Register, DHS order 2026-21447 and TechCrunch on the Bloomberg report
04 of 04
Ask an AI Who to Vote For and It Picks a Party
With the midterms on November 3, Convovo, a research studio led by Cornell professor Tapan Parikh, asked GPT and Grok through their APIs who to vote for in nearly 500 races. The prompt told them to name one candidate and not hedge. The consumer apps refused, but the APIs answered.
SpaceXAI's Grok 4.7 picked the Republican 92% of the time. OpenAI's GPT-6.1 Sol picked the Democrat 97% of the time. The two agreed less than 10% of the time. GPT picked the Democrat in 57 of 67 heavily Republican districts, and Grok picked the Republican in 47 of 72 deep blue ones. Grok leaned on X, Turning Point Action's site, and a Christian conservative voter guide. GPT leaned on Wikipedia, the AP, Axios, and Reddit. The researchers also noticed every Grok API request carried about 1,240 tokens of extra input they didn't write and couldn't see.
OpenAI's response was that a developer API with a pushy prompt is not how people use ChatGPT, which is designed to be objective by default. SpaceXAI didn't respond in time.
What I want to know is who those 1,240 hidden tokens are talking to. Look, nobody should take voting advice from a chatbot, and OpenAI has a point that the prompt was built to force an answer. But "forced to pick" is exactly how people talk to these things at 11pm the night before an election. Two models, same question, opposite answers, and both sounding sure of themselves. That tells you less about the candidates and a whole lot about who built the model.
Source: TechCrunch and Convovo's study