
OpenAI has developed an autonomous AI system capable of finding and exploiting hundreds of security vulnerabilities in widely used software.
In a new cybersecurity experiment, the company’s AI agents uncovered 948 distinct Linux kernel privilege-escalation vulnerabilities, generated working exploits, and developed patches to address the underlying flaws. The findings raise concerns about how quickly increasingly capable AI systems could identify weaknesses in software that supports millions of computers and digital services.
OpenAI disclosed these findings in a report detailing an experiment conducted in October using an upcoming cybersecurity model. The ChatGPT-maker said the entire operation cost $1.21 million in API credits and demonstrated how AI agents could automate several stages of vulnerability research that typically require specialised security expertise.
How OpenAI’s AI Agents Exploited Linux Kernel Vulnerabilities
The experiment focused on the Linux kernel, which manages the core operations of Linux-based systems, including memory, hardware resources, running applications, and communication between software and computer components. It underpins technologies ranging from Android devices and cloud infrastructure to supercomputers.
OpenAI’s research team developed an automated system that examined the kernel’s code for weaknesses that could be exploited by users with limited system permissions.
Within 12 hours, the AI agents identified 2,308 potential security flaws across selected sections of the kernel. They subsequently reproduced 1,724 vulnerabilities in isolated virtual machines, allowing the researchers to determine whether the reported weaknesses could actually be triggered.
The system then developed 1,169 working local privilege-escalation exploits and 291 proof-of-concept demonstrations involving information leaks. OpenAI subsequently conducted variant analysis and developed 948 distinct patches for the vulnerabilities behind the exploits and used this figure to establish the number of unique local privilege-escalation vulnerabilities identified during the experiment.
These flaws are concerning because local privilege escalation allows an attacker who already has limited access to a computer to obtain higher permissions. On Linux systems, this could provide root access and give the attacker extensive control over system files, applications, processes, and security settings.
AI System Generates Patches for the Discovered Vulnerabilities
Beyond identifying and exploiting the security weaknesses, OpenAI’s system also generated candidate patches to address the underlying problems in the Linux kernel.
The company incorporated patch generation into its automated research pipeline, allowing the AI agents to propose fixes alongside their vulnerability reports and exploits. However, human security researchers still reviewed the findings, assessed their severity, and examined the proposed patches before submitting them to Linux kernel maintainers.
According to OpenAI, it had already submitted hundreds of reports and patches for critical vulnerabilities, with additional submissions being coordinated through its Patch the Planet initiative.
OpenAI also reported progress by the Linux kernel community in addressing security vulnerabilities during the research period. By the time the scan and manual analysis ended after 10 days, 61 Linux privilege-escalation vulnerabilities had been patched upstream, 43 had staging patches, and another 299 were being publicly reported or discussed.
OpenAI’s Research Reveals Growing AI Cybersecurity Capabilities
The latest findings represent a significant increase in the number of vulnerabilities OpenAI’s cybersecurity models can identify and exploit.
In July 2026, an earlier experiment using GPT-5.6 Sol uncovered 166 distinct Linux kernel privilege-escalation vulnerabilities at a cost of $513,690 in API credits. The October experiment identified more than five times that number, although it cost approximately $1.21 million.
OpenAI also examined 166 exploits produced during the latest experiment and identified at least 10 previously undocumented Linux exploitation techniques. The company withheld their technical details because of the security risks involved.
The findings have prompted OpenAI to warn that attackers with access to an unpatched operating system could potentially obtain root privileges, even if they initially have access to only a low-privileged account. As such, the company recommends keeping software updated and using additional security measures, including sandboxing and isolation technologies such as Firecracker and gVisor.
As AI systems become more capable of finding vulnerabilities and developing working exploits, OpenAI believes automated security research and patch generation will become increasingly important in protecting widely used software.
