OpenAI ships GPT-6 Cyber to vetted red teams only

The best security model OpenAI has built is not one you can use. That is on purpose.

OpenAI released GPT-6 Cyber on Nov. 5, through its API and Codex, to a select group of customers in Daybreak Red, its security program for the model, according to The Decoder's account of the announcement. OpenAI says access will widen gradually to all Daybreak Red customers.

OpenAI rates its own model Critical

OpenAI rates GPT-6 Cyber "Critical" in the cybersecurity category of its Preparedness Framework, the same rating as GPT-6 Astra, per The Decoder.

The scores are OpenAI's own. On ExploitGym, a test of turning known vulnerabilities into working exploits, OpenAI reports 48.5% for GPT-6 Cyber, against 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra.

OpenAI also says the model stayed within its authorized scope in 99.93% of tests without special safety instructions, up from 94.49% for its predecessor.

Then the real-world claims. OpenAI says the model found 81 previously unknown serious vulnerabilities in Firefox, and Mozilla has accepted the reports and shipped fixes. In a Linux kernel run, OpenAI says it reproduced 1,724 vulnerabilities and built working exploits for 948, at a total compute cost of $1,213,430.

About $1,280 per exploit.

No independent lab has published tests of the model yet.

The gate matters more than the score

Yes, every big lab now has a cyber model. This one matters because of who gets it and who waits.

OpenAI says "thousands" of security professionals are in Daybreak, TechCrunch reports, and it has announced a $1 billion program for participants who protect critical infrastructure such as power grids and water systems.

Getting in is not hard. Max Justicz, CTO and co-founder of Charge Robotics, used an OpenAI model through the Daybreak Blue program to find a severe vulnerability affecting Ubuntu servers, which the Debian Project patched. He said he just uploaded a photo of his driver's license to join, per TechCrunch.

That is a light gate. It keeps the program open to small teams, and it also means the barrier is an ID check, not a background check.

Justicz wrote that he expects "some extremely crazy bugs to be found and exploited over the next few months," TechCrunch reported. Eric Wallace, a researcher on OpenAI's technical staff, said the company is "now allocating a significant amount of our frontier compute" to finding and patching vulnerabilities, per the outlet.

What it means for your code

You probably cannot use GPT-6 Cyber. The researchers auditing your dependencies soon can, and attackers will get something close from other models eventually.

The patch window is shrinking. A working exploit now costs about as much as a good laptop to build, if OpenAI's figures hold.

Two habits change today:

• Treat Firefox, Chrome and kernel updates as urgent, not routine. The fixes from this kind of hunting land in ordinary releases.
• If you run security for infrastructure or maintain a widely used open-source project, apply to Daybreak. Penetration testing at machine speed is now something you can request, not just something done to you.

Watch the expansion, not the benchmark

The next signal is how fast OpenAI widens access beyond the first Daybreak Red customers, and whether outside testers confirm that 99.93% scope figure.

Watch who gets in. Then update your browser.