OpenAI launches GPT-5.6-Cyber with reduced refusals, 95% completion on advanced cybersecurity tasks

OpenAI has introduced GPT-5.6-Cyber, a specialized model that focuses on advanced cybersecurity tasks, such as vulnerability research and exploit development. This new model is a fine-tuned version of OpenAI's most advanced general model, GPT-5.6 Sol, and has been trained specifically to improve performance on such high-risk tasks.

According to internal OpenAI benchmarks, GPT-5.6-Cyber completes 95% of advanced cybersecurity tasks, compared to 57.3% for its predecessor, GPT-5.5-Cyber, and just 1.5% for the normal GPT-5.6 Sol model.

However, GPT-5.6-Cyber is not readily available to all ChatGPT or API customers. Instead, organizations must be accepted into the Daybreak cybersecurity program, known as Daybreak Red, to access the model. This tier is intended for approved security teams conducting advanced, authorized cyber work, such as vulnerability research, penetration testing, and exploit validation.

To qualify for Daybreak Red, applicants must demonstrate a robust security program with controls like single sign-on, multifactor authentication, and a documented incident-response process.

For those who do not meet the stringent requirements for Daybreak Red, OpenAI offers the Daybreak Blue tier. This broader access tier provides approved defenders with access to OpenAI's general-purpose models, such as GPT-5.6 Sol, with adjusted safeguards for legitimate defensive work. This includes tasks like secure-code review, vulnerability discovery, malware analysis, incident response, and patch validation.

While these use cases are still sensitive, they may be more suitable for many enterprise security teams than the highly specialized GPT-5.6-Cyber model.