OpenAI has launched GPT-5.6-Cyber, a specialized cybersecurity model, to assist authorized defenders in countering increasingly sophisticated AI-driven cyber threats. The model is part of the Daybreak program, which offers two access tiers: Daybreak Blue for general-purpose models and Daybreak Red for advanced cybersecurity tasks. GPT-5.6-Cyber is designed to reduce refusals for high-risk cyber activities and improve performance on critical security workflows.
The new model is trained to enhance capabilities in areas such as zero-day vulnerability detection and exploit chain development. In internal evaluations, GPT-5.6-Cyber completed 95% of advanced cybersecurity tasks, significantly outperforming GPT-5.6 Sol, which managed only 1.5%. The model also surpassed GPT-5.5-Cyber, which completed just 57.3% of similar requests. These results were measured using an internal benchmark called the Advanced Cybersecurity Completion Rate, which tests responses to prompts involving exploit development and other advanced scenarios.
OpenAI emphasized that GPT-5.6-Cyber is tailored for use in Daybreak Red access, which is reserved for authorized security research. The model is intended to support legitimate defensive activities, such as vulnerability discovery and exploit validation, while minimizing the risk of misuse. The company also noted that early access to GPT-5.6-Cyber has been provided to trusted customer partners, who have reported significant improvements in their security workflows.
Source: openai
