OpenAI has described an experimental AI pipeline that analyzed Linux kernel findings, reproduced vulnerabilities and generated exploits that could let an unprivileged local user gain root access. In a run conducted in mid-October 2026, the system produced 948 working local privilege-escalation exploits within 30 hours, the company said.
The pipeline examined 2,308 candidate findings and dynamically reproduced 1,724 vulnerabilities. It also generated proof-of-concept exploits for 291 information-leak vulnerabilities. OpenAI said the run cost $1,213,430 in API credits, equivalent to about $1,279 per privilege-escalation exploit.
The process has five stages: identifying promising audit targets, testing candidate vulnerabilities in a sandbox, assessing whether they can be exploited, generating patches and having security researchers review the proposed exploit, patch and severity before reporting findings to maintainers.
OpenAI said the agents used at least 10 novel Linux exploitation techniques. These included methods for placing objects in kernel caches, improving exploit reliability, obtaining privilege escalation through data-only techniques and exploiting type confusion.
The work extends an earlier effort in which OpenAI said its GPT-5.5-Cyber model had discovered 24 Linux privilege-escalation exploits. The company argues that faster vulnerability discovery should be matched by automated patch generation, while emphasizing that human experts guide the process and review its output.
The report also highlights the security risks of AI tools being available to both defenders and attackers. The UK National Cyber Security Centre said AI-assisted tools can help speed up patching and reduce the time vulnerabilities remain exploitable, but warned that systems may still be at risk if the technology is not configured and used properly. It said organizations will need to adapt their patching and testing processes to keep pace with vulnerability discovery.
Sources: The Hacker News.