OpenAI has launched GPT-5.6-Cyber, a new artificial intelligence model designed to assist security professionals in identifying system vulnerabilities and constructing exploits before malicious actors can deploy similar tools.
The release expands the company’s Daybreak cybersecurity initiative with two distinct access levels. Daybreak Blue supports defensive operations such as malware analysis and incident response. Daybreak Red is intended for offensive security research, including penetration testing and exploit validation. Access to both tiers requires identity verification, account security protocols, monitoring, and legal declarations. Hardware security keys will become mandatory for all Daybreak accounts starting 1 September 2026.
High completion rates for blocked queries
GPT-5.6-Cyber is available exclusively through the Daybreak Red tier. The model is built on the GPT-5.6 Sol foundation but has been specifically trained to handle sensitive security tasks. OpenAI states that the model responds to nearly all security-related queries that other artificial intelligence systems typically reject.
An internal benchmark named Advanced Cybersecurity Completion Rate shows GPT-5.6-Cyber answers 95 percent of queries involving scenarios such as exploit chain development, authentication bypass, and privilege escalation. GPT-5.6 Sol with safety features enabled achieves a 1.5 percent completion rate. The Daybreak Blue tier reaches 2 percent. The earlier GPT-5.5-Cyber model managed 57.3 percent.
In a specific test requiring the development of a WebSocket authentication bypass for an internal admin panel, only GPT-5.6-Cyber on Daybreak Red produced functional exploit code. Every other variant refused to respond. On ExploitGym, a benchmark measuring how effectively models convert known vulnerabilities into working exploits, GPT-5.6-Cyber outperformed both GPT-5.6 Sol and GPT-5.5-Cyber.
Real-world findings
OpenAI has already applied GPT-5.6-Cyber to practical vulnerability research. The model examined V8, the JavaScript engine powering Chrome, and identified two previously unknown flaws. These issues can be linked to corrupt memory and bypass the V8 heap sandbox. Google patched the defects following a coordinated disclosure and assigned them the CVE-2026-15903 identifier.
The model also reportedly discovered at least five vulnerabilities in a popular mobile operating system. One of these is a chain of flaws allowing an application to escalate its normally restricted access rights to full administrator privileges, effectively taking control of the device. OpenAI is collaborating with Daybreak partners and the open-source community to disclose and resolve these issues.
Under the Preparedness Framework, GPT-5.6-Cyber has been rated High for cybersecurity capabilities but does not meet the Critical threshold. The recently announced Astra model is expected to reach that level. Given that GPT-5.6-Cyber is a specialised, optimised model and still falls short of Critical, the trajectory indicates AI cyber capabilities are advancing rapidly with each new generation.
What it means
Security teams now have a tool that can generate working exploit code where standard safety filters would previously block the request. This shifts the balance slightly, giving defenders a clearer view of how AI can be weaponised against specific systems. However, the requirement for hardware security keys and sandboxed environments suggests OpenAI intends to keep these capabilities strictly controlled and isolated from general use.
