OpenAI's new security model finds 948 vulnerabilities in Linux

OpenAI has developed an an autonomous cyber vulnerability discovery, exploitation, and patching pipeline that discovered a staggering 948 unique vulnerabilities in the Linux operating system. The system also created exploits to gain root access using these flaws. The entire operation was part of an automated cybersecurity research process run by OpenAI last month, costing over $1.2 million in API credits.

OpenAI's new system is a major step forward in the field of cybersecurity.

The company said it used an upcoming cyber model for the purpose. However, details about this model and its possible release date remain under wraps.

The main goal of the project was to help defenders discover and remediate vulnerabilities before they could be exploited by attackers.

Linux is a key software platform that powers Android devices, cloud infrastructure, and supercomputers.

OpenAI claims billions of people around the world use Linux-based devices.

The company's system was able to find flaws and build exploits in one go last month.

It first identified security-relevant code in the Linux kernel (over 40 million lines of code) before searching for potential vulnerabilities.

The system used a novel approach to discover vulnerabilities, employing at least 10 new Linux exploitation techniques.

OpenAI said scout agents found 5,034 security-relevant targets covering about 2.8 million lines of Linux kernel code or roughly 7% of the codebase.

Analyst agents then examined these targets and generated 2,308 candidate findings within 12 hours.

The system later reproduced each issue in isolated virtual machines and confirmed 1,724 vulnerabilities within 30 hours.

The system's exploit-development agents tested the vulnerabilities under normal security protections. They created proofs of concept for 291 information leaks and 1,169 working local privilege escalation exploits.

OpenAI said these exploits mapped to 948 unique patches, which it counted as unique local privilege escalation vulnerabilities.

Human security researchers reviewed the results before reporting these vulnerabilities and proposing fixes to Linux maintainers.

The entire operation cost $1,213,430 in API credits.

When compared with historical Linux kernel bounty rates given to users, which are $20,000 at Pwn2Own and $70,000 at TyphoonPWN, the hypothetical value of 948 vulnerabilities came out to be between $18.9 million and $66.3 million or about 16-55 times the pipeline's cost.

This shows that AI agents can find flaws much cheaper than human experts.