OpenAI has unveiled GPT-5.6-Cyber, an artificial intelligence model designed specifically for cybersecurity. The announcement, made on August 10, coincides with a reorganization of the Daybreak program, which is now divided into two access levels. The company aims to address a landscape where AI-powered automated attacks are multiplying and defenders need more precise tools.
The new model, built on the foundation of GPT-5.6 Sol, reduces rejections of dual-use requests and allows researchers to tackle tasks previously blocked by standard safeguards. However, its use will be restricted to professionals and organizations that pass a rigorous verification process.
Daybreak Blue and Red: two access levels
The Daybreak program has been divided into two categories. Daybreak Blue provides access to general-purpose models like GPT-5.6 Sol with controls tailored for defensive work, including code review, malware analysis, incident response, and patch validation. Daybreak Red, on the other hand, is reserved for specialized models like GPT-5.6-Cyber , geared towards vulnerability research, exploit validation, and offensive security testing. This separation allows OpenAI to apply stricter security measures to its most sensitive capabilities.
GPT-5.6-Cyber ​​performance: 95% of advanced requests completed
OpenAI has introduced an internal metric called Advanced Cybersecurity Completion Rate to measure the proportion of advanced requests that the system responds to. GPT-5.6-Cyber ​​completes 95% of these requests , while its predecessor, GPT-5.5-Cyber, reaches 57,3%, and GPT-5.6 Sol barely reaches 1,5% with its standard safeguards. When used through Daybreak Blue, Sol's percentage rises to 2%. These figures reflect a greater willingness to respond, although they do not guarantee the accuracy of the responses.
In ExploitGym tests, which evaluate the ability to turn known vulnerabilities into working exploits, GPT-5.6-Cyber ​​outperforms the other models compared . However, in ExploitBench, with a standard limit of 300 turns, GPT-5.6 Sol achieves the best result; increasing the limit to 600 turns reduces the difference. Furthermore, in vulnerability discovery and report writing assessments, Sol produces more detailed reports than the specialized model, which tends to be more concise.
Real findings from GPT-5.6-Cyber: vulnerabilities in Chrome and other systems
OpenAI has revealed that GPT-5.6-Cyber ​​played a role in the discovery of two vulnerabilities in Chrome's V8 engine. These flaws could be chained together to corrupt memory and escape the browser's sandbox . One of them has already been patched by Google and registered as CVE-2026-15903; the other is still undergoing coordinated disclosure. The company also states that the model has identified at least five vulnerabilities in a mobile operating system, three critical database flaws, and more than 400 privilege escalation issues in an operating system kernel, although these findings have not yet been made public.
Controlled access and program partners
Access to Daybreak is not open. OpenAI requires identity verification, usage monitoring, and contractual restrictions . Starting September 1, 2026, individual accounts will be required to use physical security keys. In addition, the company has established a partner program that includes consulting firms such as Accenture, IBM, Capgemini, EY, PwC, and NCC Group, along with security vendors such as CrowdStrike, Cisco, Sophos, Akamai, Fortinet, Cloudflare, and Palo Alto Networks. Cognizant, KPMG, and SpecterOps are also mentioned. These partners will be able to integrate Daybreak's capabilities into their products and services, but access to the underlying models is not directly transferred to end users.
The release of GPT-5.6-Cyber ​​comes after OpenAI decided to pause development of Astra, a future model that could reach a critical level of cyber-offensive capability. The company has insisted that this new model is unrelated to the Hugging Face attack, which was attributed to autonomous AI agents. In any case, the arrival of GPT-5.6-Cyber ​​presents a scenario in which security teams have more powerful tools, but also greater demands for governance and control. The reduction in rejections in the specialized model is offset by a more closely monitored operating environment, where identity, isolation, and oversight are as important as performance itself.
