OpenAI's GPT-6 Cyber Discovers 81 High-Severity Firefox Vulnerabilities, Develops Linux Exploits
OpenAI's new GPT-6 Cyber AI model has identified 81 high-severity vulnerabilities in the Firefox browser and demonstrated the capability to develop working exploits for numerous Linux kernel flaws.
OpenAI has unveiled GPT-6 Cyber, a specialized artificial intelligence model designed for cybersecurity tasks, which has already made significant contributions to vulnerability discovery. The model has identified 81 novel, high-severity vulnerabilities within the Mozilla Firefox browser. In addition to finding these flaws, GPT-6 Cyber also proposed patches for them, which Mozilla has reportedly implemented. The model also re-evaluated 30 previously known Firefox issues, concluding they were more severe than initially assessed.
Beyond browser vulnerabilities, GPT-6 Cyber has showcased its prowess in exploit development by reproducing 1,724 vulnerabilities within the Linux kernel and successfully developing working exploits for 948 of them. OpenAI stated this represents more than a five-fold increase in exploit development capability compared to its predecessor, GPT-5.6 Cyber. The model also detected critical weaknesses in Brotli, a compression library utilized by various software, including Node.js.
GPT-6 Cyber is an evolution of OpenAI's AI development, specifically trained to assist cybersecurity professionals. Unlike its more general-purpose counterpart, GPT-6 Astra, GPT-6 Cyber is designed to be less restrictive in its responses to requests from both attackers and defenders, facilitating deeper security testing and exploit development. OpenAI highlighted that while mainline models like GPT-6 Astra can perform code review and patch development, GPT-6 Cyber, available through the Daybreak Red customer program, enables more advanced exploit creation and authorized security testing.
OpenAI has integrated stronger safeguards into GPT-6 Cyber to mitigate misuse, acknowledging the increasing attempts by malicious actors to leverage AI for harmful purposes. The company emphasized that as AI models become more capable, the risks of them being used to cause harm also escalate. These safeguards are intended to help monitor for cyber-specific misalignment in AI agents, protect against stolen credentials, identify misuse, and monitor emerging threats.
In performance benchmarks, GPT-6 Cyber achieved a score of 48.5% on ExploitGym, a metric measuring a model's ability to exploit vulnerabilities, significantly outperforming GPT-5.6 Cyber's 32.1%. In terms of adherence to scope and restrictions, the new model reportedly avoided answering 99.93% of out-of-scope service requests, demonstrating dedicated training to operate within authorized parameters and report results accurately.
Early adopters of GPT-6 Cyber include prominent organizations such as Cisco and CodeRabbit. OpenAI is also expanding access through its Daybreak program, which provides select companies and researchers early access to its most capable cyber models. This initiative is part of a broader commitment, including a $1 billion pledge to a public sector defender program called Daybreak for America.
The development of specialized cyber AI models by companies like OpenAI and Anthropic (with models like Mythos and Fable) aims to equip defenders with advanced tools for identifying and patching system weaknesses. However, the accessibility of these powerful models remains a point of discussion, with some industry leaders advocating for broader access to aid a wider range of organizations in defending against sophisticated threats.