Build the Brain, Then Buy the Blame: Why Frontier AI Labs Will Enter Cybersecurity

Frontier AI labs already sell security models, and most people in security still think that's where it stops. The common view is that the labs will supply the brain and the vendors will keep selling the product, because the labs are too expensive, don't have the threat intel or the expertise, users won't trust them, and they can't beat CrowdStrike or Palo Alto Networks anyway. I think that view is wrong, and the clearest proof is the fight the labs are already having to protect their own models.

The labs are already in the security market

This year the labs stopped talking about security and started shipping it. OpenAI launched Daybreak in May to find and patch vulnerabilities, and by early September Google, Anthropic and OpenAI had all announced cyber models in the same week. Then Fortune reported that OpenAI is preparing GPT-6 Cyber together with its first product built to deploy it, which would make it the company's fourth cyber model this year.

So the labs are in, and the real question is whether they stay suppliers to security vendors or become security vendors themselves. To answer that, I'd start with the thing they fear most.

Why do frontier AI labs need security at all?

Open-weight models are the biggest threat to US frontier labs. Every week we hear them called the enemy and a national security risk, and that's true in the way that matters most for the business, because if open models get stronger and cheaper to run, customers switch away from Anthropic and OpenAI. The labs know this, which is why prices keep dropping.

The way open models catch up is distillation, which means copying a big model by studying its answers. Someone sends it huge numbers of questions through thousands of fake accounts, collects the answers and trains their own model on them. OpenAI took this all the way to Congress in February, and the fight has only grown since.