OpenAI releases GPT-6 Cyber to vetted Daybreak Red customers

What OpenAI actually shipped

GPT-6 Cyber is a cyber-specialised variant of GPT-6 Astra. OpenAI fine-tuned it on crash reproduction, codebase vulnerability search and working exploit development.

The company rated it Critical under its Preparedness Framework, one step above the High rating it gave GPT-5.6 Cyber. OpenAI says the label reflects attack capability, not perfect reliability. OpenAI defines Critical as a tool-augmented model able to develop functional zero-day exploits in many hardened real-world systems without human intervention. Launch access runs through the API and Codex to a select group of Daybreak Red customers, gated by identity and eligibility checks. Organisations must run single sign-on, multi-factor authentication, managed devices, least-privilege access and default API IP allowlisting. OpenAI says it will expand access progressively to all Daybreak Red customers.

The numbers OpenAI is leading with

On ExploitGym, GPT-6 Cyber scored 48.5%, against 32.1% for GPT-5.6 Cyber, a 51% relative gain.

In the Linux kernel study, an autonomous pipeline analysed 2,308 candidate issues, reproduced 1,724 of them and produced working local privilege escalation exploits for 948. Il Sole 24 Ore put the API cost at about $1.2 million, or $1,279 per exploit, with the runs conducted on lab virtual machines. OpenAI reported 304 critical findings to Linux maintainers alongside suggested fixes. On Mozilla's Firefox, the model found 81 previously unknown vulnerabilities, all rated high severity by Mozilla, and fixes for many have shipped.

OpenAI released all of those figures; Il Sole 24 Ore noted that no independent analysis had confirmed them.

Where the evidence runs out

Irregular ran an external evaluation of the model.

In that FrontierCyber test, GPT-6 Cyber solved 90 of 288 challenges against 96 for GPT-6 Astra, and neither model compromised a fully hardened target. The system card reports a 42% mean arbitrary-code-execution success rate on recently disclosed V8 and Chrome vulnerabilities, while warning that reliable exploitation remains hard. In the ExploitGym Honeypot, which plants an out-of-scope service beside the authorised target, GPT-6 Cyber exploited it in 0.07% of graded rollouts without safety instructions, against 5.51% for GPT-5.6 Cyber; OpenAI says it avoided out-of-scope services in 100% of tests with deployment safety prompts.

Why the timing is the argument

OpenAI's own framing is about speed. "It is clear that we have a limited window to use frontier models to secure software," the company said.

The UK's AI Security Institute put the capability gap between open and closed frontier models at four to seven months, down from six to ten for most of 2025. Vals AI's CyberOffense benchmark sets agents loose on simulated corporate networks with no human help. Deepseek V4.1 Flash reached 72.8% of objectives against 80.8% for GPT-6 Sol, at about $3 per run versus roughly $22. A cyber-range run with Deepseek V4-Pro cost $1.19, against $85 for Anthropic's Opus 4.5, according to Vals AI. Vals researchers say that once a model's weights are public, no provider can control how it is used.

What happens next

From 29 November 2026, OpenAI plans to use content from Daybreak Blue sessions on personal accounts to improve its models, including safeguards against cyber misuse.

The Hacker News reported that the plan applies even when the "improve the model for everyone" setting is off. OpenAI also says it will widen Daybreak Blue alpha access for GPT-6 Astra to more organisations and individual security researchers. It plans to bring Daybreak tools to organisations of all sizes, independent researchers and open-source maintainers. Neither a price for wider access nor a general availability date has been published.