Busy week in AI, and not the fun kind. OpenAI built a model so good at breaking into software that it rated it at its own highest danger level. Corporate AI spending basically went flat for two months. And a giant Wyoming data center changed hands because the builder could not get it built. Let's get into it.
01 of 03
OpenAI's GPT-6 Cyber Is Locked Behind a Velvet Rope
OpenAI launched GPT-6 Cyber on November 5. It is not a public release. Only approved customers in Daybreak Red, OpenAI's program for authorized penetration testing and exploit development, can use it, with access widening to all approved Daybreak Red customers in the coming weeks.
The numbers are the story. On ExploitGym, a benchmark where the model has to actually exploit a specified flaw, GPT-6 Cyber scored 48.5%. The older GPT-5.6 Cyber scored 32.1%, and OpenAI's own flagship general model, GPT-6 Astra, scored 42.4%. OpenAI rates GPT-6 Cyber "Critical" under its Preparedness Framework for finding and exploiting software vulnerabilities, which is the top of its scale.
OpenAI also says the model chewed through 948 Linux kernel vulnerabilities and helped prep patches, a bug hunt TechCrunch reports cost about $1.2 million in compute. OpenAI researchers separately found 81 previously unknown high-severity bugs in Firefox, and Mozilla has started fixing them.
The guardrails: IP allowlisting by default, cyber-specific misuse monitoring, and new blocks on high-risk activity aimed at critical infrastructure like power plants and water systems. In a test built around staying inside an authorized target, the model avoided out-of-scope services 99.93% of the time with no safety prompt and 100% with one. Its predecessor wandered out of bounds in about 5.5% of runs.
Here is where I land on it. Gating this thing is the right call, and I am glad somebody is pointing frontier compute at Linux patches pro bono. But the defenders TechCrunch talked to have a fair gripe. Max Justicz, CTO of Charge Robotics, found a severe Ubuntu server bug (CVE-2026-109518) in about two hours with OpenAI's defender program, then said that program got "nerfed." Meanwhile bad actors are running open-weight models with the guardrails stripped off. If the good guys get the velvet rope and the bad guys get the back door, the rope is not doing much.
Sources: eSecurity Planet, TechCrunch
02 of 03
Ramp Says Business AI Spending Just Hit the Brakes
Ramp, the corporate card company, tracks what businesses actually pay for AI, and its latest data is not what the IPO crowd wants to see. Token spend grew 42% month over month in July. Then it grew just 1.5% across August and September combined. October picked back up, with spend up 8.9% from September 23 to October 23.
The split between companies is wild. The median firm spends $12 to $15 per employee per month on AI. The top 10% spend $600 to $1,000. The top 1%, the ones Ramp calls "AI-pilled," peaked around $10,000 per employee per month this summer and slipped a little below that in August and September.
Ara Kharazian, Ramp's lead economist, put it bluntly: "Almost the entire spend on AI is from venture-backed, high-growth tech companies." When that top 1% slows down, it is enough to stall revenue even while usage keeps climbing. Part of the reason is price cuts. OpenAI halved API prices for GPT-6 Sol and GPT-6 Luna in September, and Anthropic cut prices on Fable 5.1 and Opus 5.5, which a company filing said knocked about $600 million off its revenue growth.
The timing stings. Bloomberg reports Anthropic's potential IPO valuation fell from $2 trillion in early October to somewhere between $1.2 and $1.3 trillion. OpenAI lowered its own target to $1.25 trillion from $1.4 trillion.
What I want to know is who the next buyer is. If the whole AI economy is a few hundred hot startups buying tokens from each other, that is not a market, that is a potluck. Cheaper tokens are great for people like me building stuff. But the labs priced their valuations like every dentist office in Lubbock was going to be spending a grand per employee. The median company is spending fifteen bucks. That gap is the whole ballgame.
Sources: Ramp AI Index, TechCrunch
03 of 03
Crusoe Hands Its Wyoming Mega Campus to Google
Crusoe, the Denver data center developer recently valued at $30.9 billion, confirmed Friday it is selling its unfinished campus in southeastern Wyoming to Google. Crusoe announced the project last year with Tallgrass, a Blackstone-backed infrastructure company, and TechCrunch reports it was planned to scale to 10 gigawatts of power.
Google did not disclose the price, but it says it plans to invest $17 billion to develop the site. The awkward part: Crusoe had tried and failed to land Google as a customer for this exact project. Bloomberg reported back in June that Google raised concerns about cost and construction timeline under Crusoe's management. Getting equipment and power, plus growing local pushback, made it harder.
Crusoe is not hurting for work. It raised $3.9 billion in September, is building another campus for Google near Amarillo, built the Abilene site for Oracle and OpenAI, has a 900 MW campus going for Microsoft, and reportedly signed a $13 billion, five-year cloud deal with Jane Street.
My read is simple. This is what it looks like when the hard part of AI turns out to be dirt, steel, and transformers instead of math. A developer worth thirty billion dollars could not get a customer comfortable with its timeline, so the customer just bought the site. Expect more of this. The hyperscalers have the cash, and they are done waiting on middlemen. And with Meta already warning that election backlash against data centers could make AI expansion harder, owning the land outright starts looking less like a flex and more like insurance.
Sources: Crusoe, TechCrunch