OpenAI releases GPT-5.6-Cyber, a new model aimed at cybersecurity workflows that include zero-day vulnerability discovery and exploit-chain development. The Next Web reports the company had previously paused a model over cyber risk before shipping this version on Monday, and says GPT-5.6-Cyber is trained to “refuse fewer” higher-risk dual-use cyber requests.
According to Help Net Security, the model is built on GPT-5.6 Sol and is designed to improve performance on certain exploit development and advanced security research tasks, with a reduced refusal rate for security researchers’ requests in higher-risk categories. Both outlets describe OpenAI’s stated goal as improving utility for legitimate cybersecurity work while still applying safeguards.
Access to the model is limited: both reports say availability runs through Daybreak, a vetted program for cybersecurity professionals, specifically referencing Daybreak Red as the higher tier where the model can be used.
