OpenAI announced the expansion of the Daybreak cybersecurity program, introducing two levels of access and releasing the GPT-5.6-Cyber model specifically trained for cybersecurity. The core logic behind this initiative is that threat actors will increasingly use AI to launch cyberattacks at unprecedented speed and scale, and the window for defense is rapidly narrowing. Therefore, it is essential to equip trusted defenders with cutting-edge intelligence before advanced AI attack capabilities are widely deployed.

Daybreak Blue offers the cutting-edge general model GPT-5.6 Sol and removes system-level safeguards that hinder legitimate defensive work, making it suitable for most security professionals; Daybreak Red provides specialized cybersecurity models tailored for high-risk tasks such as authorized vulnerability research and exploit validation.

image.png

The GPT-5.6-Cyber is the core of this release, built on the GPT-5.6 Sol and specifically trained for tasks such as zero-day vulnerability discovery and exploit chain development, while reducing the rejection rate for high-risk dual-use cybersecurity requests. In an internal advanced cybersecurity completion rate evaluation conducted by OpenAI, the GPT-5.6-Cyber completed 95% of advanced cybersecurity requests, compared to only 1.5% for the GPT-5.6 Sol and 57.3% for its predecessor, the GPT-5.5-Cyber. Jared Atkinson, CTO of SpecterOps, stated that the model completed work that previous models had failed to solve over weeks of intermittent efforts in a single day.

Practical Results: From Chrome Engine to Mobile Operating Systems

The capabilities of GPT-5.6-Cyber are not limited to benchmark tests but have also uncovered numerous high-risk vulnerabilities in real software. OpenAI used the model to study the V8 JavaScript engine of the Chrome browser and discovered two previously unknown vulnerabilities that could be exploited to break memory and escape the V8 heap sandbox. After verification, the research team reported them through a coordinated disclosure process to Google, which fixed them and assigned them the identifier CVE-2026-15903, a high-severity vulnerability. The root cause was an error in the V8 optimizer compiler that skipped a security check when converting values to integers. In addition to the V8 vulnerability, GPT-5.6-Cyber also found at least five vulnerabilities in mainstream mobile operating systems and built an exploit chain from untrusted applications to local privilege escalation. It discovered three critical vulnerabilities in popular databases, including paths for remote code execution, and more than 400 vulnerabilities in operating system kernels that could lead to privilege escalation.

OpenAI is working closely with Daybreak partners and the open-source community to disclose and fix these vulnerabilities in mobile operating systems, databases, and kernels. In terms of security assessments, GPT-5.6-Cyber outperformed the GPT-5.6 Sol and GPT-5.5-Cyber in the ExploitGym exploit evaluation, showing clear advantages in zero-day vulnerability discovery and severity calibration. However, the model slightly lagged behind the GPT-5.6 Sol in the vulnerability report writing evaluation due to its concise output. OpenAI will continue to optimize it in future versions.

Emphasizing Tiered Control and Security Safeguards

Lowering security safeguards means taking risks beyond standard usage, but OpenAI believes that equipping defenders with cutting-edge intelligence is crucial for accelerating cybersecurity defenses. Access to Daybreak is controlled by multiple mechanisms, including identity verification, account security monitoring, approved usage restrictions, and legal statements. Since September 1st, all personal accounts must now use hardware security keys as a mandatory requirement. OpenAI also strongly advises customers using Codex to switch from full access mode to automatic review mode, which evaluates and intercepts actions with significant destructive risk before executing operations requiring elevated privileges.

As GPT-5.6-Cyber pushes cybersecurity capabilities to new heights, the AI-driven arms race in offensive and defensive automation has inevitably begun in full force. Whoever achieves a better balance between security and capability will gain the upper hand in the next round of cyber space competition.