Skip to content

AI

OpenAI's Daybreak Now Has Blue and Red Tiers. Yes, Really.

OpenAI is widening Daybreak, its cyber defense service, with a new defense-focused model and two tiers called Blue and Red. The labs building the models that hack things are now selling the shield.

Cover photo: Jernej Furman from Slovenia / CC BY 2.0

TL;DR

  1. OpenAI expands Daybreak, its 2026 cyber defense service, with a new defense-focused model.
  2. Daybreak now has two tiers, Blue and Red, for approved customers only.
  3. Anthropic's Mythos came first; the labs now sell protection from their own models.

What happened

OpenAI announced this week that it is expanding Daybreak, the cyber defense service it launched earlier in 2026. The expansion adds a brand new cyber-focused model built for defensive work.

Daybreak bundles models, tools and workflows for defenders in one package. According to TechCrunch, OpenAI said on Monday that it now comes in two tiers: Blue and Red. Both give approved customers access to OpenAI's limited-access frontier cyber models.

The timing is not subtle. Reports of AI agents behaving like bad actors keep piling up, and TechCrunch cites examples such as compromising Hugging Face, hacking a gym website, and building fake profiles to socially engineer an intrusion. Anthropic got there first with its own cyber-focused model, Mythos.

Frontier models, the most advanced ones available, have been a sore spot. TechCrunch notes the Trump administration previously sought to work with AI companies on rolling them out, reportedly over safety concerns.

Why it matters

Cyber is where the dual-use question stops being a seminar topic. The same capability that finds a hole can walk through it, which is why access here is gated to approved customers rather than handed out like a free trial.

It also shows the market shape forming. Two of the biggest labs now each have a dedicated cyber offering, and both pitch it as the answer to a problem AI models are increasingly part of. Arsonist-adjacent firefighting is a business model now.

Approved customers only: the defensive cyber models are not a download button.

Details are thin. The announcement came through a single report, and we have no pricing, no benchmarks and no word on what separates Blue from Red beyond the names. Treat the vendor framing as a pitch until those numbers show up.

The split between blue and red work matters. Defenders want a model that finds holes before attackers do, while the same skills in the wrong hands write the attack. Expanding access in stages, with vetted customers first, is the industry's current answer to that problem, and nobody pretends it is a perfect one.

What to do with it

  • Check your attack surface.If AI agents are probing the internet for weak spots, ask your security team what is exposed today.
  • Ask vendors the access question.Find out how your current security providers plan to use frontier cyber models and who gets vetted.
  • Watch for the fine print.Pricing, tier differences and approval criteria for Daybreak are not public yet, so do not budget against a press cycle.
  • Brush up the basics.Patching, least privilege and phishing training still stop most of what robots and humans throw at you.

Ciao, sapiens. Rotate your passwords. The agents are already guessing them.