OpenAI launches GPT-6 Cyber to identify and fix software vulnerabilities

A new tool for cyber defense

OpenAI has introduced GPT-6 Cyber, a specialized artificial intelligence model built to identify software flaws and assist experts in repairing critical vulnerabilities. The company aims to use this technology to help defenders stay ahead of malicious actors who seek to exploit system weaknesses. This release represents a targeted effort by the firm to lead in the emerging field of AI-assisted cybersecurity.

The model is designed to perform tasks such as vulnerability discovery and exploit testing. By automating parts of the research process, it allows security professionals to scan large amounts of code more efficiently. OpenAI positions this as a way to make advanced cybersecurity research more accessible to those tasked with protecting digital infrastructure.

This new release puts OpenAI in direct competition with other major tech players. The company is vying for market share against Anthropic’s Claude Mythos Preview, Google’s Gemini 3.8 Flash Cyber, and Microsoft’s AI-focused MDASH system. While these products differ in specific features, they all target the same space of software protection and automated bug finding.

Performance benchmarks and results

OpenAI claims that GPT-6 Cyber represents a significant leap forward from its predecessor, GPT-5.6 Cyber. In testing on the ExploitGym benchmark, which measures a model's ability to exploit known vulnerabilities, GPT-6 Cyber achieved a score of 48.5 percent. In comparison, the older GPT-5.6 Cyber model scored only 32.1 percent on the same metric.

The practical applications of this increased performance are already visible in recent research. OpenAI researchers used the frontier AI to develop working exploits for 948 different Linux kernel vulnerabilities. These results demonstrate the model's ability to navigate complex system code to find and utilize specific weaknesses.

The model also successfully identified 81 previously unknown high-severity vulnerabilities within the Mozilla Firefox web browser. After these flaws were discovered, Mozilla was notified of the findings. The browser company has already begun the process of releasing fixes for several of the critical issues identified by the AI.

Comparing the different models

OpenAI has provided specific comparisons to show how GPT-6 Cyber stands out against its own internal lineup. The company stated that GPT-6 Cyber outperforms GPT-6 Astra on most cyber tasks. This remains true even in instances where the Astra model does not refuse to perform the requested task.

The distinction between these models is important for understanding how OpenAI manages risk. By creating a specialized

Strategic competition and progress

The release of GPT-6 Cyber marks a deliberate move by OpenAI to reclaim its position in the global AI race. By focusing on specialized cyber tasks, the company is attempting to stay ahead of rivals who are also developing tools for vulnerability discovery and software protection. This move highlights a shift toward high-stakes, practical applications of frontier models.

OpenAI asserts that the new model shows strong progress over its predecessor, GPT-5.6, specifically in the areas of finding and exploiting software flaws. The company notes that GPT-6 Cyber is a major step up from the previous version and even outperforms GPT-6 Astra on most cyber tasks. This remains true even in scenarios where the Astra model does not refuse to perform the requested request.

Controlled access for security

Because the ability to find critical vulnerabilities carries a high risk of misuse by hackers, OpenAI is not releasing GPT-6 Cyber to the general public. Instead, the company is making the model available only through its API and Codex to a specific group of customers. These users must be approved for the Daybreak Red program to gain access to the tool.

The primary goal of this restricted rollout is to ensure that advanced cybersecurity research remains in the hands of trusted defenders. By limiting who can interact with the model, OpenAI hopes to facilitate the discovery of bugs while preventing unauthorized attacks. This balance aims to provide security experts with powerful tools without creating a blueprint for malicious actors to exploit systems.