OpenAI has unveiled a novel AI model, GPT-5.6-Cyber, designed specifically for cybersecurity applications. Announced on Monday, the model focuses on enhancing tasks such as vulnerability research, penetration testing, and incident response. Built on the capabilities of GPT-5.6 Sol, this version aims to reduce refusals for complex cybersecurity tasks, including zero-day vulnerability detection and exploit chain development.
Introducing Daybreak Red Access
The introduction of GPT-5.6-Cyber comes with the launch of a new access tier called Daybreak Red. This tier allows firms to utilize OpenAI’s specialized cybersecurity models for tasks such as authorized vulnerability research and security testing. The model is an evolution of GPT-5.5-Cyber, initially released in June 2026, and offers more permissive features for cybersecurity professionals.
OpenAI has established an internal metric known as the Advanced Cybersecurity Completion Rate to assess the model’s performance. This metric evaluates the model’s ability to perform tasks related to exploit-chain development and other advanced security scenarios. Results indicate that GPT-5.6-Cyber completes 95% of these requests, a significant improvement over its predecessors.
Improved Security Capabilities
GPT-5.6-Cyber demonstrates enhanced performance in cybersecurity workflows, particularly in exploit development and advanced security research. The ExploitGym benchmark shows that the model surpasses both GPT-5.6 Sol and GPT-5.5-Cyber. Despite these advancements, the model occasionally produces less detailed vulnerability reports, which can impact its performance in open-ended tasks.
Notably, the model has identified several high-severity vulnerabilities, including CVE-2026-15903, a critical flaw in the V8 JavaScript engine. This vulnerability, which allows potential code execution, was patched by Google in July 2026. The model also flagged over 400 vulnerabilities that could lead to privilege escalation in various systems.
AI’s Role in Cybersecurity
OpenAI’s new model is available to select partners such as Accenture and Cisco, aiming to enhance their cybersecurity measures. These models are crucial as cybercriminals increasingly leverage AI to execute attacks more efficiently. However, while AI tools like GPT-5.6-Cyber significantly aid in identifying vulnerabilities, they still require human expertise for effective patching without introducing new issues.
Research indicates that AI-generated patches only resolved vulnerabilities without altering application behavior 26% of the time. This highlights the need for continued human oversight in cybersecurity. OpenAI emphasizes that despite the risks associated with reduced safeguards in AI models, providing access to advanced intelligence is vital for improving and automating cyber defenses.
