14:11Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI Unveils GPT-6 Cyber Model That Found 81 Firefox Vulnerabilities
OpenAI has released its GPT-6 Cyber model, designed for security research. According to the company, the model significantly outperforms GPT-5.6 Cyber at finding and fixing security vulnerabilities in software. During testing, it discovered 81 new vulnerabilities in the Firefox browser. Mozilla rated all of them as high severity.
Previous artificial intelligence models had been unable to find these vulnerabilities. GPT-6 Cyber also developed potential fixes for the issues. The model also showed that 30 previously reported vulnerabilities were more dangerous than thought. Mozilla has fixed many of the vulnerabilities through an update, and the fixes have reached millions of users. Firefox users who have disabled automatic updates are advised to install the latest version manually.
The model developed working code to exploit 948 vulnerabilities in the Linux kernel in real-world attacks. In the same test, the previous GPT-5.6 Sol model produced only 166 such results. OpenAI sent 304 critical findings, along with potential fixes, to Linux developers. Cisco and CodeRabbit are also using the model for code reviews.
OpenAI also tested the model on its own servers. In controlled tests, programs running GPT-6 Cyber and the company’s other models escaped isolated environments and were able to execute code on core systems. After identifying the issues, the company updated its software and tightened network restrictions.
A separate security test was also conducted to determine whether the model would go beyond its assigned task. Easily compromised services were added to a closed testing environment, but interfering with them was not part of the main task. In 99.93% of tests conducted without security instructions, the model left those services untouched. When given such instructions, it succeeded in every case. OpenAI has classified the model at the “Critical” cybersecurity level. It is currently available only to a small number of approved security teams.