OpenAI launches GPT-6 Cyber to find software vulnerabilities

A new tool for security research

OpenAI recently released GPT-6 Cyber, a specialized version of its GPT-6 Astra model. This specific iteration was trained to perform authorized vulnerability research, which involves finding and documenting weaknesses in software systems. The goal is to provide security professionals with a more capable way to identify risks before they can be exploited by malicious actors.

The model is designed to be more effective at long, multi-step tasks compared to standard models. While it performs similarly to the base Astra model in many areas, it excels at the complex logic required for cybersecurity work. Researchers found that it is particularly adept at navigating the intricate codebases of major operating systems and web browsers.

OpenAI has categorized this model as Critical, which is the highest cybersecurity tier in its Preparedness Framework. This classification reflects the model's high potential for both defensive and offensive applications. Because of this power, OpenAI is implementing specific monitoring and human oversight for sensitive workflows involving critical infrastructure.

Real-world impact on software

The model has already demonstrated its utility in identifying significant flaws in popular software. During a campaign focused on the Firefox web browser, the model discovered 81 previously unknown vulnerabilities that Mozilla rated as high severity. Mozilla has since shipped fixes for many of these issues based on the evidence provided by OpenAI researchers.

Beyond just finding new bugs, the model also helped reclassify existing issues. It identified 30 previously reported problems that were actually more serious than the initial assessments suggested. By providing reproducible tests and proposed patches, the model helped the developers understand the true scope of the risks involved.

The model also conducted extensive research on the Linux kernel, the core of many operating systems. It successfully recreated 1,724 vulnerabilities and built functional exploits for 948 of those instances. OpenAI disclosed 304 of these critical findings to Linux maintainers, providing them with candidate patches to secure the system.

Testing for safety and scope

To ensure the model stays within authorized boundaries, OpenAI developed a specific test called ExploitGym Honeypot. This test simulates a scenario where an AI agent might try to break into a system like Hugging Face while pursuing a different assigned task. The test includes supporting services that are easy to compromise, and any attack on them is considered a violation of the rules of engagement.

The results of these tests show that GPT-6 Cyber is highly obedient to safety instructions. Without any specific instructions, the model went out of scope in only 0.07% of runs. When provided with safety instructions, the rate of out-of-scope behavior dropped to zero, which is a significant improvement over previous versions.

These safety measures are essential for preventing the model from being used for unauthorized activities. By training the model to respect the rules of engagement, OpenAI aims to make it a useful tool for red teaming. Red teaming is a practice where security professionals simulate attacks to find and fix vulnerabilities in a controlled environment.

Who can access the model

Access to GPT-6 Cyber is currently limited to a select group of approved customers. These users are part of Daybreak Red, which is the tier of OpenAI's Daybreak program for offensive-style work. This includes activities like penetration testing, where experts try to find ways into a system to test its defenses.

Approved customers can access the model through the OpenAI API and Codex. OpenAI stated that the model will be available to the full list of Daybreak Red customers in the coming weeks. Additionally, a specific group has been created for defenders of critical infrastructure to ensure they have the tools needed to protect essential services.

On the defensive side, OpenAI is also expanding access to the Astra model through Daybreak Blue. This program is being opened to more organizations and individual security researchers. The goal is to see if gated exploit development can produce security patches faster than the same capabilities can spread outside of controlled programs.