- GPT-6 Cyber can identify software vulnerabilities and develop exploits.
- The model scored significantly higher than its predecessor in the ExploitGym test.
- Researchers discovered multiple high-severity vulnerabilities using GPT-6 Cyber.
- The model excelled in adhering to authorized testing environments, achieving a compliance rate of 100%.
OpenAI has launched a specialized cybersecurity model named GPT-6 Cyber, which is capable of identifying software vulnerabilities and developing effective exploits, while also demonstrating improved capabilities within authorized security testing parameters. The new model scored 48.5% in the ExploitGym test, a benchmark for assessing a model's ability to exploit specified vulnerabilities against remote targets, marking a 51% increase from the 32.1% score of GPT-5.6 Cyber. OpenAI stated that GPT-6 Cyber outperformed its general-purpose GPT-6 Astra model in most cybersecurity tasks, even though Astra does not reject requests.
The company has upgraded the cybersecurity classification of GPT-6 Cyber to "Critical" under its readiness framework, compared to "High" for its predecessor. This classification reflects the model's enhanced capabilities in understanding software, identifying vulnerabilities, and developing effective exploits.
Application of the Model in Vulnerability Research
In a vulnerability research initiative targeting the Mozilla Firefox browser, the model's capabilities were significantly demonstrated. OpenAI researchers utilized GPT-6 Cyber to discover 81 previously unknown vulnerabilities that Mozilla classified as high severity. The model also identified 30 previously reported issues that were more severe than initially assessed. Researchers provided Mozilla with evidence, reproducible tests, and suggested remediation strategies to assist its engineers in addressing these vulnerabilities.
Mozilla subsequently began releasing fixes for several reported vulnerabilities.
Development of New Assessment Methods
OpenAI has also developed a new assessment method for conducting offensive network penetration tests, which requires the model to probe web applications without access to the source code. GPT-6 Cyber achieved a vulnerability recall rate of 40.3%, while GPT-6 Astra scored 32%. Unlike traditional vulnerability assessments, which examine source code, network penetration testing requires GPT-6 Cyber to identify weaknesses by directly interacting with running web applications, testing the model's ability to discover exploitable entry points based solely on the behavior and responses exposed by the running services.
The model also demonstrated its ability to reproduce and exploit vulnerabilities in the Linux kernel. It reproduced 1,724 vulnerabilities and developed effective exploits for 948 of them. OpenAI disclosed 304 critical findings to Linux maintainers, along with candidate remediation strategies. The company had previously detailed this research separately.
Beyond exploit development, OpenAI focused on ensuring that GPT-6 Cyber respects the constraints of authorized security assessments. In its ExploitGym Honeypot evaluation, researchers placed the model in an isolated environment containing vulnerable supporting services. While exploiting these services could reveal indicators of the task, doing so exceeded authorized boundaries. GPT-6 Cyber avoided exploiting out-of-scope services in 99.93% of tests without the need for security prompts.
With prompts enabled, its compliance rate reached 100%. GPT-5.6 Cyber scored 94.49% without prompts enabled. OpenAI stated that specialized training helps the model adhere to researchers' constraints, remain within authorized testing environments, and accurately report findings. The company combines these capabilities with additional safeguards, including network-specific monitoring, restrictions on high-risk activities involving critical infrastructure, and human oversight of sensitive workflows.
GPT-6 Cyber is initially available through API and Codex to a small group of approved Daybreak Red customers. OpenAI plans to expand access to other Daybreak Red customers while also providing them with GPT-5.6 Cyber.
Breakthrough Advances of GPT-6 Cyber in Cybersecurity
As cyberattacks continue to rise, enhancing vulnerability identification capabilities has become crucial. The launch of GPT-6 Cyber not only demonstrates significant advancements in identifying and exploiting vulnerabilities but also showcases its compliance with authorized security testing. These capabilities greatly increase the model's potential applications in the cybersecurity field, providing stronger protections for businesses and organizations. OpenAI's research not only contributes to improved security but also enhances the efficiency of vulnerability remediation, having a profound impact on the overall cybersecurity ecosystem.

