Three stories ran the AI world this past week, and none of them were about a chatbot getting a little smarter. One is about the most important CEO in the industry, one is about voters, and one is about a model OpenAI itself rates as dangerous. Let's get into it.
01 of 03
Nvidia's Board Is Investigating Jensen Huang
On Sunday, Nvidia's board announced an independent investigation into CEO Jensen Huang. It will be run by the independent directors with outside legal counsel.
Here is how we got here. On November 6, the New York Times reported on a memo, reportedly circulated among US policymakers earlier this year and said to be based on US intelligence. The memo claims China's Ministry of State Security recruited an executive at Megaspeed, a Singapore-based company, to spy on Huang. The Financial Times then reported the memo also claimed Huang had a romantic relationship with that executive.
Huang then sent a message to the company admitting an affair. "I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did. I am deeply sorry," he wrote, according to TechCrunch. The board says it has "no reason to believe" national security or company secrets were compromised, and that it has "confidence in Jensen and his leadership team." Earlier, an Nvidia spokesperson told the Times that Huang's interactions with the executive were "not unusual" and that Nvidia does not share classified information with customers.
To be clear about what is confirmed: the affair is confirmed by Huang himself. The spying claim is not. It lives in a memo nobody has published, and no US official has confirmed it on the record.
Here is where I land on it. The affair is his family's business, not mine. The spy part is everybody's business, because this is the guy who sits in the middle of every chip export fight between Washington and Beijing. "No reason to believe" is a company grading its own homework. I want to see what the outside counsel actually finds, and I want it published, not summarized in a two line press release.
Sources: TechCrunch on the board investigation, TechCrunch on the intelligence memo
02 of 03
Data Centers Went on the Ballot and Got Whipped
Last Tuesday's midterms had a sleeper issue, and it was data centers. According to the Christian Science Monitor, there were 41 referendums, initiatives, and referrals on local ballots across 10 states. In all but one, data center development got constrained or banned outright.
The breakdown: at least 18 bans, 4 moratoriums, and 10 new zoning style restrictions. The rest were opposition votes without a hard restriction attached. The lone exception was a North Dakota town of under 900 people that declined to require extra review. Ohio led everybody with 21 measures.
Voters went after the people, too. Piqua, Ohio recalled three city commissioners. Urbana, Ohio recalled its mayor. Both cities banned data centers the same day. Yukon, Oklahoma's mayor lost a recall, and so did a council member in Mason, Michigan.
The industry noticed. Meta president Dina Powell McCormick said at the WSJ Tech conference that the Democratic wins "are going to make it a lot harder for companies to build data centers," and admitted the industry has "not done as good a job as we needed to explain the benefits." Meta has data centers in 23 states. Meanwhile Crusoe, the Denver developer valued at $30.9 billion, confirmed it sold its unfinished southeastern Wyoming project to Google. That site was pitched as scaling to 10 gigawatts. Google did not disclose the price, but says it plans to invest $17 billion building out the campus.
What I want to know is whether anybody in Big Tech is actually listening. Folks are not voting against "AI." They are voting against their power bill, their water, and a humming warehouse next to the high school. "We didn't explain the benefits" is the wrong lesson. When 40 out of 41 ballot measures go against you, the pitch is not the problem. The deal is.
Sources: Christian Science Monitor, TechCrunch on Meta, TechCrunch on Crusoe and Google
03 of 03
OpenAI Built a Hacking Model It Rates "Critical"
On November 5, OpenAI released GPT-6 Cyber, a model built to find and exploit security flaws. OpenAI rated it "Critical" for cybersecurity under its own Preparedness Framework, a notch above the "High" it gave the previous GPT-5.6 Cyber.
The numbers, as reported by SecurityWeek: on ExploitGym, a test of whether an AI can exploit a known flaw in a remote target, GPT-6 Cyber scored 48.5 percent, up from 32.1 percent for GPT-5.6 Cyber. It found 81 new Firefox flaws that Mozilla rated high severity, and Mozilla has started shipping patches. On the Linux kernel, it reproduced 1,724 vulnerabilities, and OpenAI reported 304 critical issues to maintainers with proposed fixes. In a test where it was handed easy-to-hack decoy services that were off limits, it stayed away from them 99.93 percent of the time without safety prompts, and never touched them with the prompts on.
You can't just sign up. Access is limited to select approved customers of Daybreak Red, OpenAI's program for advanced exploit work and authorized testing, through the API and Codex. OpenAI says it will roll it out to all Daybreak Red customers over time. Industrial security firm Dragos helped test the safeguards, and Dragos's Rob T. Lee put it bluntly: "The skill floor is dropping for the reconnaissance and vulnerability research needed before an attack."
Not everybody is sold. Max Justicz, CTO of Charge Robotics, told TechCrunch he found a severe Debian-related bug using an earlier OpenAI model for about $5,000, and argued defenders are at a disadvantage because attackers can just grab open-weight models with the guardrails stripped off.
I think gating this behind a vetted program is the right call, and 81 real Firefox bugs getting patched is a real win for regular people. But I keep coming back to Justicz's point. "Thousands" of approved defenders is a nice number until you compare it to everybody else on the internet. The bad guys don't fill out an application. The good news is only good if the patches ship faster than the open models catch up.
Sources: OpenAI, SecurityWeek, TechCrunch