OpenAI published research on autonomous vulnerability discovery, demonstrating their AI’s capability by identifying 948 valid local privilege escalation vulnerabilities in the Linux kernel.
According to the report, OpenAI’s prototype initially found 2,308 candidate vulnerabilities in the Linux kernel, dynamically reproduced 1,724 of those vulnerabilities, and ultimately confirmed 948 local privilege escalation exploits that would allow unprivileged local users to gain root access. The researchers also generated candidate patches for maintainer review.
The discovery process was computationally expensive, consuming $1.2 million USD in API credits to identify the vulnerabilities.
One notable finding from OpenAI’s research concerns the relative performance of different AI models. Their GPT-5.5 Cyber model found more than 5 times more novel local privilege escalation vulnerabilities than their GPT-5.6 Sol model, despite the fact that three months of hardening and patches had been applied to the Linux kernel between the two scans.
OpenAI researchers noted broader implications of their work, stating: “Within the next few years, we believe that if there are vulnerabilities in a given software project, frontier models will likely be able to discover them and even less capable models may be able to develop exploits once disclosed.”
The researchers highlighted potential responses to this emerging capability, including building and maintaining high-assurance software—particularly for critical systems like operating systems, browsers, image parsing libraries, and TCP/IP stacks. They identified promising areas for future research, including training models to write secure code by default, minimizing attack surface by removing unnecessary features, and formal verification as approaches to effectively scale software security alongside finding and fixing existing vulnerabilities.
Key facts
- OpenAI’s AI found 948 confirmed local privilege escalation vulnerabilities in the Linux kernel
- The prototype initially identified 2,308 candidate vulnerabilities and reproduced 1,724 of them
- The vulnerability discovery cost $1.2 million in API credits
- OpenAI’s GPT-5.5 Cyber model found 5x more novel vulnerabilities than the GPT-5.6 Sol model despite intervening kernel patches