OpenAI has revealed that its latest AI model, developed for security research, successfully created working exploits for 948 vulnerabilities in the Linux kernel, the foundational code for countless systems worldwide.
What Did OpenAI's AI Find?
In a landmark study, OpenAI deployed
a new specialised model, GPT-6 Cyber, to audit the Linux kernel, a massive and critical open-source project that powers everything from Android phones to the majority of the world's cloud servers. The AI didn't just find theoretical flaws; it developed 948 distinct, working exploits for what are known as local privilege escalation (LPE) vulnerabilities. In simple terms, these are flaws that would allow an attacker who already has basic access to a machine to gain full 'root' control, effectively taking over the entire system. The research, which cost over $1.2 million in API credits, involved the AI model scanning roughly 7% of the kernel's code, or about 2.8 million lines. The model first identified over 2,300 potential issues, then worked to confirm them, and finally proved their severity by creating the exploits.
A Leap in AI Capability
The model used, GPT-6 Cyber, represents a significant advance over its predecessors. In a similar test, a previous version, GPT-5.6 Sol, had produced exploits for 166 vulnerabilities. The new model developed working exploits for more than five times that number, showcasing a dramatic increase in capability. This research was not just about finding flaws; it was also about fixing them. OpenAI's system generated candidate patches for the vulnerabilities it found, which were then reviewed by human experts before being disclosed to Linux maintainers. The work also uncovered 81 previously unknown, high-severity vulnerabilities in the Firefox browser, many of which have since been patched by Mozilla. This demonstrates the AI's power as a tool for defenders to find and fix holes before they can be widely abused.
The Cybersecurity Double-Edged Sword
The stunning success of this research highlights a critical dilemma in the world of AI: the 'dual-use' problem. A tool powerful enough to find and fix nearly a thousand security flaws is also powerful enough to be weaponized by malicious actors. If AI can automate the discovery of vulnerabilities at this scale, it dramatically lowers the barrier for cybercriminals to find and exploit zero-day flaws—vulnerabilities that are unknown to those who should be mitigating them. OpenAI is acutely aware of this risk. The research was conducted in a controlled environment, and access to these powerful cyber-focused models is heavily restricted through gated programs like 'Daybreak Red', which are available only to approved security researchers and organizations for defensive purposes. The company is also investing heavily in safeguards to prevent misuse, training its models to refuse harmful requests and monitoring for malicious activity.
The Future of AI_Powered Security
This research signals a new era in cybersecurity. The sheer scale of modern software, like the 40ndillionndline Linux kernel, makes it impossible for human experts to manually scrutinize every part. AI offers a way to analyse code at a scale and speed that was previously unimaginable, offering a powerful advantage to defenders. The goal is to move from a reactive security posture to a predictive one, finding and fixing flaws before they ever become a problem. However, this creates an arms race. As defenders get better tools, attackers will inevitably try to leverage the same technology. OpenAI notes that within a few years, it's likely that if a vulnerability exists, a powerful AI will be able to find it. This makes the development of secure software from the ground up more critical than ever.
