OpenAI announced GPT‑6 Cyber on 5 November for authorised security work. The model can look for vulnerabilities, test whether they are exploitable and help develop candidate fixes. Access to advanced testing is restricted through Daybreak Red.
Where the improvement shows
On ExploitGym, a controlled test in which models are given a vulnerability to exploit, GPT‑6 Cyber scored 48.5%, up from 32.1% for GPT‑5.6 Cyber. OpenAI classifies its cybersecurity capability as Critical, a designation under the company’s own framework. The test measures performance on a defined set of tasks and does not tell teams how often the model will succeed on their own systems.
The same ability can support defence or be misused. Access controls and rules of engagement therefore matter alongside detection rates. Security teams still need to review findings, reproduce them in approved environments and test proposed fixes before deploying them.
