OpenAI Unveils GPT-6 Cyber With Major Leap in AI-Powered Cybersecurity
OpenAI has introduced GPT-6 Cyber, a specialised AI model for cybersecurity research. It has helped uncover 81 Firefox vulnerabilities, develop 948 Linux kernel exploits and identify a critical Brotli security flaw.
GPT-6 Cyber is an upcoming version of AI from the company OpenAI that aims at assisting security analysts in discovering software vulnerabilities, developing exploits and fixing the bugs that can be later exploited by the hackers. This model marks a notable evolution of AI-based cybersecurity as the company emphasizes its superior performance in detecting software bugs, exploiting them, and taking into consideration security constraints when conducting research.
The thing is that the new AI-based model is specifically tailored to deal with cybersecurity challenges and can perform such operations as vulnerability discovery, exploiting of the discovered vulnerabilities, code analysis and web penetration testing. According to the representatives of the company, the AI is already actively helping security analysts to investigate software projects such as Linux kernel and Mozilla Firefox for their vulnerabilities.
One of the major accomplishments was achieved in the area of Linux kernel security research. As explained by OpenAI, GPT-6 Cyber managed to reproduce 1,724 vulnerabilities and create effective local privilege escalation exploits for 948 of them. The exploits may have allowed a regular user to escalate their privileges. OpenAI also revealed 304 high-severity findings to Linux maintainers along with corresponding patches while the process of solving other vulnerabilities is still in progress.
GPT-6 Cyber was able to make considerable contributions in the area of Firefox security research as well. OpenAI researchers found out using GPT-6 Cyber about 81 previously unknown high-severity vulnerabilities in Mozilla's Firefox browser. Furthermore, they were able to find out that another 30 vulnerabilities had been previously known but turned out to be even more severe. OpenAI notified Mozilla of these vulnerabilities and offered them patches.
The second noteworthy finding pertains to Brotli, a widely-used library for compression purposes. Researchers from OpenAI made use of the new GPT-6 Cyber model along with other OpenAI tools to create exploits for the highly vulnerable system referred to as CVE-2026-111497. The mentioned vulnerability may enable attackers to run remote code in some cases. Its CVSS 4.0 rate is estimated at 9.2, putting the vulnerability in the "Critical" category.
At the same time, OpenAI emphasizes considerable progress in the performance of the model in exploiting vulnerabilities. For instance, on ExploitGym (benchmarking tool assessing whether AI models can utilize known vulnerabilities to develop effective exploits), the GPT-6 Cyber model managed to score 48.5%. Its previous version (GPT-5.6 Cyber) scored 32.1% there. Hence, a gain in 16.4 percentage points can be observed.
But enhanced cybersecurity features come with their own set of dangers especially if such cybersecurity technologies get into the wrong hands. According to OpenAI, GPT-6 Cyber has been designed to adhere strictly to the boundaries of the authorized security research. In the alignment test, GPT-6 Cyber adhered to the boundaries of the authorized scope 99.93% of the time in comparison to GPT-5.6 Cyber’s 94.49%. And when deployed safety prompts were engaged, GPT-6 Cyber was able to adhere to the boundaries of the authorized scope 100% of the time.
Because of the advanced nature of GPT-6 Cyber and because of the possible dangers it may pose to cybersecurity, OpenAI has categorized it as Critical under its Preparedness Framework as opposed to GPT-5.6 Cyber which is categorized as High in terms of cybersecurity. Additional safeguards have been put in place to ensure no harmful activities are conducted involving critical infrastructures, industrial control systems, essential services, and election systems.
GPT-6 Cyber is not currently available for the general public. The first stage will involve providing it via OpenAI’s API and Codex to approved Daybreak Red customers. This will be due to the fact that the software is intended for use in high-level and approved cybersecurity tasks like penetration testing, red teaming, and vulnerability research. OpenAI will be expanding access to other approved customers of Daybreak Red, and at the same time improving its security measures.
GPT-6 Cyber is an example of how AI can become a much more powerful weapon in defence of software and digital infrastructures. It will be able to detect previously unknown vulnerabilities, create functioning exploits, and assist cybersecurity teams to come up with a patch. Nevertheless, as AI-based cyber tools continue to evolve and their capabilities continue to grow, it becomes evident how important restricted access and safeguards are.
Comments