OpenAI launches GPT-6 Cyber to find software vulnerabilities

A tool for proactive defense

OpenAI has introduced GPT-6 Cyber, a specialized artificial intelligence model designed to assist security researchers in identifying and fixing software vulnerabilities. The company classified this model as "Critical" within its internal safety framework for cybersecurity. It aims to help researchers find flaws and test how they might be exploited so that patches can be created before bad actors find them.

The release comes as a direct response to the growing trend of people using AI to conduct cyberattacks. By providing a tool that can discover weaknesses systematically, OpenAI hopes to give security teams a head start. The model is intended to support a wide range of users, from large global enterprises to independent software maintainers who need to keep their code secure.

This new model is part of a broader initiative called Daybreak. This program provides security research tools and offers subsidies to help frontline defenders access the technology. OpenAI wants to ensure that the people responsible for defending digital infrastructure have the best possible resources at their disposal.

Performance and real-world impact

To prove the model's effectiveness, OpenAI reported results from independent benchmarks. On the ExploitGym test, which measures how well an AI can exploit specific vulnerabilities, GPT-6 Cyber reached a score of 48.5%. This is a significant jump from its predecessor, GPT-5.6 Cyber, which scored 32.1% on the same test.

The model has already been put to work on real-world software systems. For instance, researchers used it to develop 948 privilege escalation exploits in Linux. These are types of attacks that allow a user to gain higher-level permissions than they were originally granted on a computer system.

In another notable test, the model identified 81 previously unknown high-severity vulnerabilities in the Mozilla Firefox web browser. OpenAI shared these findings with Mozilla, who then released patches to fix many of the issues. This shows how the tool can be used to secure popular software used by millions of people every day.

Who can use the technology

While OpenAI plans for broader availability in the future, access to GPT-6 Cyber is currently limited. Only a group of vetted Daybreak Red customers can use the model right now. These users can access the tool through the company's API and Codex platforms.

Several major organizations are already utilizing the technology for their security operations. Companies such as Cisco, Calif, and CodeRabbit are using the model for code reviews and vulnerability research. These teams use the AI to scan their software for potential holes that could be exploited by hackers.

OpenAI researchers have also used the model to protect their own internal infrastructure. They conducted red-teaming exercises, which are simulated attacks designed to find weaknesses. These tests led to the creation of software patches and the implementation of tighter security controls within OpenAI's own systems.

Safety and alignment measures

Because the model is so capable at finding vulnerabilities, OpenAI has introduced strict safeguards to prevent it from being used for harm. They have categorized the model's capabilities as high-risk. To manage this, they implemented misalignment monitoring to ensure the AI follows specific instructions without deviating into harmful behavior.

The company also uses IP allowlisting to control which specific internet addresses can access the model. Additionally, they have built in protections against high-risk cyber assistance. These layers of security are meant to ensure that the tool remains in the hands of legitimate researchers rather than malicious actors.

During alignment evaluations, the new model stayed within its authorized task boundaries in 99.93% of tests without specific deployment prompts. When users provided specific prompts to guide the tasks, the model stayed within those boundaries 100% of the time. These numbers suggest that the model is highly responsive to the specific limits set by the human operators.

The broader business context

The launch of GPT-6 Cyber occurs during a period of significant change for OpenAI. Earlier this month, the company lowered its target for its current funding round by $10 billion. This move happened as the technology sector saw a general downturn in stock prices and increased investor anxiety regarding AI spending.

Despite these financial pressures, OpenAI continues to push forward with specialized models like GPT-6 Cyber. The company noted that as models become more capable, there are more attempts to use them to cause harm. This creates a constant cycle of development where the company must build better defenses as the technology itself becomes more powerful.

The goal of the Daybreak initiative is to provide a sustainable way for organizations to stay ahead of cyber threats. By subsidizing access for frontline defenders, OpenAI is attempting to create a standard for how AI can be used responsibly in cybersecurity. The success of this model will likely determine how much of the security industry moves toward AI-driven defense.