Reading Room

OpenAI has released a new version of its AI model built just for cybersecurity work, called GPT-6 Cyber, and it is only letting a small group of approved security customers use it for now. That gating choice matters because the same skills that make an AI good at defending computer systems can also make it good at attacking them.

GPT-6 Cyber beats OpenAI's general purpose model, GPT-6 Astra, on most cybersecurity tasks, according to OpenAI. The company says the risk that the model could be used to carry out cyberattacks if it ended up in the wrong hands is why access is restricted rather than opened to everyone, the way most new OpenAI models are. Fortune first reported in September that OpenAI was building a specialized cyber model off GPT-6, so this release confirms what had been rumored.

This is not a one-off. GPT-6 Cyber is OpenAI's fourth cybersecurity-focused model this year, after versions released in April, June, and August. OpenAI has also hired a new chief revenue officer, Dali Rajic, who previously worked in cybersecurity software. Taken together, the pattern reads less like a safety side project and more like OpenAI building a real sales line aimed at security teams and the big business customers that employ them.

The money behind this is large and growing fast. The research firm Gartner forecast in September that worldwide spending on AI for cybersecurity would reach $51.3 billion this year, nearly double last year's $25.9 billion. The reasoning is straightforward: as AI models get better at finding software flaws and turning them into working attacks, the businesses on the other side come under more pressure to strengthen their own defenses, and that pressure turns into budget. OpenAI is leaning into that budget as an important path to growing its sales to big business customers.

What is worth watching from here is the access model itself, not just the product. A specialized, high-capability tool kept behind a small approved customer list, rather than released broadly, is likely to become a template other AI companies reach for whenever a capability cuts both ways: useful for defense, dangerous for offense. If that pattern holds, expect more handle-with-care releases like this one, not fewer.

Anthropic has reportedly signed a $3.4 billion deal for gas fired power (electricity made by burning natural gas) to run its data centers, according to a Yahoo Finance report. It is a sign of how much raw energy AI companies now need just to keep scaling up, a cost that eventually shows up in electricity prices and in fights over new power plants.

An AI system built on an Anthropic model reportedly helped settle a math conjecture that had gone unsolved for almost 20 years, with implications for what is called quantum speedup (how much faster a quantum computer can solve a problem than a normal one), according to Quantum Zeitgeist. It is an early, concrete sign of AI doing real research work, not just answering questions, worth watching as the technology keeps improving.