OpenAI agents discover over 400 Linux root exploits in hours
OpenAI presented results showing AI agents identified over 400 zero-day vulnerabilities allowing local user to root escalation on Linux systems. The agents produced working exploits in an average of two and a half hours, far outpacing human reverse engineers.
At an OpenAI security conference, the company presented results from an experiment where AI agents were tasked with escalating privileges on Linux systems. The agents discovered over 400 zero-day vulnerabilities allowing local user to root escalation. Each exploit produced a binary that instantly granted root access. On average, it took the agents about two and a half hours to research, identify, test, validate, and produce the deliverable for each exploit.
This speed contrasts sharply with human reverse engineers, who may take weeks to identify the same exploits or understand the breadcrumb trails leading to them. The agents also identified new methods and technologies that had never been published, demonstrating the ability to connect disparate, non-connected pieces of information, such as S3 bucket contents and login pages, at a scale and speed far beyond human capability.
Caleb Sima stated he was previously wrong to believe AI could not make these disparate connections, noting that the speed and innovative methods of the agents are surprising and alarming. The findings suggest that AI agents can rapidly identify and exploit complex security flaws, posing a significant challenge to traditional defensive strategies that rely on slower human analysis cycles.