New OpenAI AI finds 948 vulnerabilities in the Linux kernel
OpenAI has discovered 948 previously unknown security vulnerabilities in the Linux kernel with a new AI model for IT security. The vulnerabilities could allow attackers with local access to gain complete control of a system.
Multiple agents on duty
For the investigation, the company relied on an automated sequence of specialized AI agents that analyzed thousands of potential errors and developed numerous working exploits within a few days shared OpenAI with. The Linux kernel forms the foundation of numerous systems, including Android devices, cloud infrastructure, and supercomputers. Although the source code contains more than 40 million lines, the team focused on around 2.8 million lines, around seven percent of the total.
Instead of having a single AI agent search for security vulnerabilities, a multi-stage pipeline was developed. First, so-called scout agents identified accessible code areas that can also address unprivileged users. This resulted in 5034 investigation targets. Analyst agents found 2,308 possible vulnerabilities within twelve hours. Triage agents then checked them in isolated virtual machines. After another 30 hours, 1724 errors were successfully reproduced.
The AI systems also generated 291 proof-of-concepts for information leaks and 1,169 working local privilege escalation exploits. After comparing overlapping results, 948 independent vulnerabilities remained that actually pose some risk. According to OpenAI, the exploits also worked under a standard Linux desktop configuration with protection mechanisms such as KASLR, SMEP and SMAP enabled. At least ten previously unknown attack techniques were used.
First bugs already fixed
OpenAI has already submitted hundreds of bug reports and suggested fixes to Linux developers. During the ten-day research project, 61 privilege escalation vulnerabilities have already been fixed, and provisional patches exist for 43. Another 299 cases are being discussed publicly. Microsoft could also benefit from the new technology: Since the company has access to OpenAI’s current models, it is conceivable that they are already being used to examine the Windows source code. However, it is not known whether future security updates will actually contain relevant AI findings.
RS News or Research Snipers focuses on technology news with a special focus on mobile technology, tech companies, and the latest trends in the technology industry. RS news has vast experience in covering the latest stories in technology. Can be reached at author2@researchsnipers.com
