OpenAI has released the GPT-5.6-Cyber model, enhancing vulnerability discovery and security research capabilities.

August 11th, OpenAI announced the expansion of the cybersecurity defense program Daybreak and the launch of the GPT-5.6-Cyber model dedicated to the cybersecurity field. The model aims to assist authorized security researchers and defense teams in enhancing their vulnerability discovery, threat analysis, and security testing capabilities. As attackers increasingly leverage AI to launch faster and larger-scale cyber attacks, defense teams need to proactively access advanced AI capabilities. Daybreak offers two types of access: Daybreak Blue, aimed at most defense teams, provides the GPT-5.6 Sol and other general models for vulnerability discovery, secure code reviews, malware analysis, incident response, and patch validation; Daybreak Red targets advanced security research, offering the GPT-5.6-Cyber for authorized vulnerability research, vulnerability validation, and security testing.

The GPT-5.6-Cyber, based on the GPT-5.6 Sol, is tailored for cybersecurity tasks, including zero-day vulnerability discovery and vulnerability exploitation chain analysis. It has been used in real-world vulnerability research and has identified critical vulnerabilities in software, including the Chrome V8 JavaScript engine. Furthermore, OpenAI stated that GPT-5.6-Cyber has also aided in discovering high-risk vulnerabilities in various domains, such as privilege escalation flaws in mobile operating systems, remote code execution vulnerabilities in databases, and hundreds of privilege escalation vulnerabilities in operating system kernels.

OpenAI also emphasized that Daybreak Red will only be open to approved individuals and organizations, with access controls through measures such as authentication, account security, monitoring, usage restrictions, and legal disclaimers. The company stated that it will continue to enhance security monitoring, access management, and model security testing to mitigate the risk of misuse of advanced cybersecurity models.