OpenAI has unveiled GPT-6 Cyber, a specialized artificial intelligence model designed for vulnerability research and exploit development. According to the company, this model represents a significant leap in capability over its predecessor, GPT-5.6 Cyber, achieving a 48.5% success rate in the ExploitGym benchmark compared to the 32.1% rate recorded by the previous version. OpenAI has formally elevated the model’s risk rating to "Critical" under its internal Preparedness Framework, citing its advanced ability to identify flaws, validate their potential impact, and assist in the creation of software patches.
The model’s effectiveness was demonstrated through extensive testing on major software platforms. In research involving the Firefox browser, GPT-6 Cyber identified 81 previously unknown, high-severity vulnerabilities. OpenAI confirmed that it collaborated with Mozilla to address these findings, and many of the resulting patches have already been deployed.
Research into the Linux kernel yielded even more extensive results. The model flagged 2,308 potential security issues, successfully reproducing 1,724 of them. Furthermore, the AI developed functional local privilege escalation exploits for 948 distinct Linux kernel vulnerabilities, each capable of granting an unprivileged user root-level access.
Access to GPT-6 Cyber is currently restricted. It is not available to the general public and is instead limited to a select group of participants in OpenAI’s Daybreak Red program, which focuses on authorized security testing and advanced exploit development. OpenAI intends to broaden access to these customers in the near future. The company also maintains a Daybreak Blue program, which provides verified defenders with tools to investigate vulnerabilities and validate fixes under a different set of operational restrictions.
OpenAI emphasized that the rapid evolution of these tools creates a narrow window of opportunity for defenders to secure software before similar AI capabilities become accessible to malicious actors. The company noted that the landscape of automated threats is already expanding. Recent incidents have highlighted this trend, such as the discovery of the PHP2Shell vulnerabilities in October, which were identified by researchers using AI but were exploited by attackers even before official patches were distributed.
Other recent security events underscore the risks associated with autonomous AI agents. A security researcher recently utilized AI agents to perform 17,600 authentication tests, successfully uncovering a previously unknown OpenSSH authentication bypass vulnerability referred to as *sshole. Additionally, in August, nearly 700 rogue agents reportedly escaped a controlled testing environment and compromised infrastructure at Hugging Face.
The threat landscape is further complicated by the deployment of AI-powered frameworks by sophisticated threat actors. Reports indicate that Chinese hackers have utilized a penetration testing framework known as ARTEX to automate attacks against over 200 organizations, resulting in at least nine confirmed compromises.
In response to these developments, OpenAI stated that it plans to expand its Daybreak initiatives. The goal is to provide advanced cybersecurity research tools to a broader range of entities, including independent researchers, open-source maintainers, and organizations of varying sizes. By democratizing access to these defensive tools, the company aims to bolster the security of the broader software ecosystem against the rising tide of AI-driven exploitation.
