OpenAI has launched a new AI model targeting professionals in network security.

The company stated that this new model, named GPT-6 Cyber, performs better than the general-purpose GPT-6 Astra in most network security tasks. Due to concerns that this model could be used to launch cyberattacks if it falls into the wrong hands, OpenAI is currently only available to a select number of approved security customers.

Fortune reported first in September that OpenAI plans to release a model based on GPT-6 that is specifically aimed at network security.

As the AI model becomes increasingly capable of detecting software vulnerabilities and converting them into executable attacks, businesses are facing growing pressure to strengthen their network defenses. Gartner predicted in September that global AI cybersecurity spending this year will reach $51.3 billion, nearly double the $25.9 billion spent last year.

OpenAI is increasingly betting on this market and viewing it as an important path to drive corporate sales growth. This is the fourth cybersecurity-oriented model launched by OpenAI this year, with previous versions released in April, June, and August respectively. The company has also hired Dali Rajic, who comes from a background in cybersecurity software, as its Chief Revenue Officer.

Google and Anthropic are also competing for secure customers. Anthropic's Glasswing project allows audited users to utilize capabilities that might otherwise be restricted due to security measures. Google previously emphasized a lightweight, low-cost model for repeatedly scanning large codebases; now, through its Fairwind project, it also enables audited defenders to access its cutting-edge Gemini 4 Argon models without any network security barriers.

OpenAI's GPT-6 and Cyber focus on advanced vulnerability research and vulnerability exploitation development. In order to test server defenses, researchers may attempt to invade servers to observe what attackers can do after discovering vulnerabilities. The same methods can also be used for malicious attacks; therefore, general AI models impose restrictions, which may also hinder legitimate security work. OpenAI indicates that companies train their Cyber models both to enhance specialized cybersecurity skills and to reduce rejections of tasks that could potentially assist either attackers or defenders.

OpenAI has tested GPT-6 Cyber on the web browser Firefox of Mozilla as well as on the Linux kernel. Firefox must be capable of handling web pages that may contain malicious content, while ensuring that user data is not exposed; meanwhile, the Linux kernel controls access to machine resources. Both are open-source projects, and their code has been reviewed, tested, and strengthened by developers and security researchers for many years. However, the AI model is still able to identify new vulnerabilities.

According to OpenAI, researchers utilized GPT-6 and Cyber to discover 81 previously unknown vulnerabilities with a high severity level, and they developed functional exploit code for 948 kernel vulnerabilities, including 588 vulnerabilities that were previously unknown. The exploit code for these Linux vulnerabilities could enable attackers with ordinary local access rights to obtain root privileges, which represents the highest level of control over the system. The number of exploit codes for Linux vulnerabilities developed by the researchers is more than five times that of a study in July that used GPT-5.6 and Sol; GPT-5.6 and Sol were considered top models in network security tasks when they were released that month.

OpenAI currently provides two levels of access rights to trusted cybersecurity professionals through the Daybreak project: Daybreak Red supports advanced tasks such as simulating attacks and building tools for red team exercises, while Daybreak Blue focuses on defensive tasks such as code review, malware analysis, and incident investigation.

OpenAI indicates that GPT-6 and Cyber will initially be available only to a small group of approved Daybreak Red customers, and will be gradually extended to all Red customers over the coming weeks. At the same time, OpenAI also states that they are expanding the early testing of Astra for defensive security work to more organizational and individual security researchers through Daybreak Blue.