Barely a day goes by without news of an AI stepping out of line. Just this week a Claude agent hacked a gym’s website, and before that it was Hugging Face’s datasets. So the labs whose models are causing the trouble are expanding their security offerings. On Monday, OpenAI extended its cyber-defense program Daybreak, which it had launched earlier this year, not long after Anthropic introduced its cyber-focused model Mythos.
Two tiers: Blue and Red
Daybreak now comes in two tiers. Both give approved customers access to OpenAI’s frontier cyber models, which are otherwise available only in limited form.
Blue is the more restrained option. It covers classic defense: incident response, malware analysis, patch validation. OpenAI calls it the “recommended starting point for most defenders” — enough, in other words, for the average company.
Red goes further. Here users get purpose-trained security models for penetration testing and vulnerability research. And Red comes with the new model, GPT-5.6-Cyber, which only exists at this tier. It’s built on GPT-5.6 Sol and tuned for certain security tasks. For now, access is limited to a handful of trusted partners, including Accenture, IBM, CrowdStrike, and Cloudflare.
The marketing question
You have to read all this with a second eye. Critics point out that these threat warnings double as a sales pitch for the labs. Whoever knows the risks first also sells the protection most credibly. OpenAI puts it this way: attackers will increasingly use AI for their strikes, at growing speed and sometimes fully autonomously, so the window for defenders to prepare is narrowing.
My take
What interests me most here is the symmetry. With GPT-5.6-Cyber, OpenAI is basically building the counterpart to Anthropic’s Mythos — two labs, the same idea, almost in lockstep. Both are saying: our models are good enough to be dangerous, so we’re handing them out to the good guys under control.
That’s both logical and unsettling. Logical, because the same capability that enables an attack also strengthens the defense. Unsettling, because the race between offense and defense is now being run inside models, not people anymore. And the companies driving both sides of that race are the same ones. My advice: this is a spot where looking closely pays off more than being reassured.
Sources: