OpenAI Releases GPT-6 Cyber to Select Security Teams

OpenAI launched GPT-6 Cyber on November 5, a model trained to find software flaws, write working exploits and suggest the patches.

You can’t use it yet. It’s going to select customers of Daybreak Red, OpenAI’s program for authorized defensive security work.

So why should you care if you run a website? OpenAI says it’s better at finding holes in web apps without seeing their code than its standard model, and your site is a web app.

What OpenAI says it can do

Here are OpenAI’s own numbers, as reported by The Decoder and The Verge:

• ExploitGym, a test of whether a model can exploit a known flaw in a remote system: 48.5%, up from 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra, OpenAI’s flagship.
• Web apps with no source code to read: about 40%, against 32% for GPT-6 Astra.
• Firefox: 81 flaws nobody knew about. Mozilla has shipped the fixes.
• Linux kernel: 1,724 vulnerabilities reproduced and working exploits built for 948, more than five times what GPT-5.6 Sol managed.

The kernel run cost about $1.2 million in compute, or roughly $1,300 per exploit.

That said, these are OpenAI’s numbers, and nobody has published independent tests yet.

Who gets it, and how OpenAI fences it in

OpenAI rates the model Critical for cybersecurity under its Preparedness Framework, meaning it can find and build working zero-day exploits in many hardened systems with no human help.

GPT-5.6 Cyber was rated High.

Does it stay in bounds? In a honeypot test based on the time OpenAI’s agents breached Hugging Face, it strayed less than 1% of the time, down from about 5%.

Here’s who gets what:

• GPT-6 Cyber: select Daybreak Red customers through the API and Codex now, and all of them in the coming weeks.
• GPT-5.6 Cyber: every Daybreak Red customer.
• GPT-6 Astra: alpha access to Daybreak Blue, the lower tier, widens to more organizations and individual security researchers.

Daybreak Blue is easier to reach. One founder told TechCrunch he joined it with a photo of his driver’s license, and that he’s spent about $5,000 on tokens and plans.

Then, with GPT-5.6 Sol, he found a severe flaw that could let hackers take over Ubuntu servers in two hours. Debian has patched it.

Why it matters

You won’t get GPT-6 Cyber. Neither will most people who’d attack your site.

But that isn’t the comfort it sounds like. Security pros, TechCrunch reports, say bad actors can run open-weight models with the guardrails stripped out.

And Epoch AI researchers showed an open GLM model could have exploited known, unpatched Firefox flaws.

My read: finding a flaw in your stack is getting cheap. OpenAI’s kernel run worked out to about $1,300 per exploit, where bug bounties have paid $20,000 to $70,000 for comparable finds.

So the slow part is you.

A hole in your plugin, theme or PHP version gets found faster now, but it’s only closed on your site once you, your developer or your host installs the update.

I’d do three things this week:

• Turn on automatic updates for WordPress core and the plugins you trust, and delete the ones you don’t use.
• Ask your host which PHP version your site runs. The PHP2Shell flaws showed you’re only as patched as your server.
• Work through your WordPress security checklist.

Still running a plugin that hasn’t seen an update since 2024? That’s the one I’d remove first.

Common Questions

Can I use OpenAI’s GPT-6 Cyber to scan my own website for vulnerabilities?

Only if you’re one of the select Daybreak Red customers who get GPT-6 Cyber through the API and Codex now, or one of the rest of the Daybreak Red customers who get it in the coming weeks.

What is OpenAI’s Daybreak Red program?

Daybreak Red is the top tier of Daybreak, OpenAI’s program for authorized defensive security work, and select Daybreak Red customers get GPT-6 Cyber first. Daybreak Blue, the lower tier, gives security defenders access to frontier models such as GPT-5.6 Sol, which one founder used to find a severe flaw that could let hackers take over Ubuntu servers.

Has anyone tested GPT-6 Cyber independently?

Not yet. Every number about GPT-6 Cyber, including its 48.5% on ExploitGym and the working exploits built for 948 Linux kernel flaws, is OpenAI’s own, and nobody has published independent tests yet.