In the spring, shortly after Anthropic launched Project Glasswing around its Claude Mythos model, OpenAI started its own program called Daybreak. In August it expanded it. If you've seen the name and weren't sure what it is, here's the plain version.
What Daybreak is
Daybreak is how OpenAI gives cybersecurity teams access to its most capable models and tools for defense. It started with a cyber-tuned model, Codex Security (an agent that scans codebases, confirms findings, and helps write patches), and partner projects like one with Trail of Bits to find and fix bugs in open source software.
It's not a product you sign up for with a credit card. OpenAI controls access through identity verification, monitoring, restrictions on how it's used, and legal agreements.
Blue and Red
In August, OpenAI split Daybreak into two tiers, according to CNBC and TechRadar:
- Daybreak Blue gives access to OpenAI's general models, like GPT-5.6 Sol, with safeguards loosened for defensive work: finding vulnerabilities, reviewing code, analyzing malware, incident response, and checking patches. OpenAI calls it the starting point for most organizations.
- Daybreak Red goes further, with models trained specifically for security, for authorized vulnerability research, exploit validation, and security testing.
The new model in Red is GPT-5.6-Cyber. OpenAI says it completes 95% of the specialized requests security researchers send it, up from 57% for the previous cyber model. The regular GPT-5.6 Sol, by comparison, completes only 1.5% of those same requests because its safeguards block them.
That number tells you something. GPT-5.6-Cyber is built on GPT-5.6 Sol. The underlying capability is similar. The big difference is what each one is allowed to do, and for whom.
Then in September, alongside its GPT-6 Astra launch, OpenAI announced a $1 billion effort to bring these tools to defenders of essential services like critical infrastructure.
Why you're not in it
If you're a five-person SaaS company, you're almost certainly not getting into Daybreak Red. That's by design. These programs are aimed at big security teams, critical infrastructure, and open source maintainers, and that's a reasonable place to start.
But think about what happens next.
The biggest companies and the most important software get hardened first, with the best tools. Banks, cloud providers, operating systems, major open source projects. That's good for everyone.
It also means attackers who want an easy win will find fewer of them at the top. So they'll look further down, at companies that don't have a security team, don't patch quickly, and have never had a pen test. With AI tools making attacks cheaper to run at scale, going after a thousand small targets is more practical than it used to be.
What that means for you
You don't need Daybreak. You need to not be the softest target on the list. The basics still do most of the work:
- Patch quickly, especially anything with a CVE attached.
- Use the public models on your own code. The general models from OpenAI and Anthropic will review your code for security problems now. You don't need special access for that.
- Lock down access. MFA everywhere, least privilege, no shared admin accounts.
- Watch your logs. Know what normal looks like so you notice when something isn't.
- Test from the outside at least once a year and after big releases.
The best security tools right now are going to the biggest companies first. The attackers have already noticed where that leaves everyone else.
Start with the free scan. If you want the outside view, the $495 pen test comes back in 24 hours.