OpenAI has made GPT-6 Cyber public as its most capable cybersecurity model. The model found previously unknown high-severity vulnerabilities in Firefox and the Linux kernel, and OpenAI said it outperformed GPT-6 Astra on most cybersecurity tasks.
OpenAI’s newest cyber model
OpenAI introduced GPT-5.6 Cyber in August with its Daybreak Red and Blue access tiers for security researchers. Since then, the company has released GPT-6 Astra as its flagship model and the lower-priced GPT-6 Sol and Luna models. GPT-6 Cyber now joins that line of models as an internal system turned public.
The model is trained for security work
GPT-6 Cyber is trained to find vulnerabilities, build working exploits, and help developers with fixes. OpenAI said the model is already being used to help defend security-sensitive software such as the Linux kernel, Firefox, and other projects. Greg Brockman said the company invested heavily in trust and trained the model to stay within the scope set for it.
Benchmark results and vulnerability findings
On ExploitGym, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. That is a 51% relative improvement. On the new RemoteExploitBench benchmark, the model completed 95% of exploit development tasks.
In Firefox, GPT-6 Cyber found 81 new high-severity vulnerabilities and identified 30 previously known issues that were more serious than originally reported. Mozilla has already released fixes for many of those vulnerabilities. OpenAI also classified GPT-6 Cyber as having “Critical” cybersecurity capabilities under its Preparedness Framework.
Availability and planned expansion
GPT-6 Cyber is available first through the API and Codex to select approved Daybreak Red customers. Broader access is planned. GPT-5.6 Cyber is now available to all Daybreak Red customers, and OpenAI is expanding Daybreak Blue access to GPT-6 Astra for additional organizations and individual security researchers.
OpenAI is also expected to launch GPT-6.2 Astra and GPT-6.2 Sol later this week. The company said those models will have meaningful improvements over their predecessors. Along with Azure and AWS, OpenAI’s Frontier models will also be available via Google Cloud on Day 1.
Licensed under CC BY-ND 4.0 International