OpenAI indicates that its autonomous security agent discovered 948 different vulnerabilities within the Linux kernel, developed usable exploit code, and drafted a repair plan for review – the entire process cost approximately $1.2 million in API credits.

In the results released on Wednesday, OpenAI stated: 'We chose Linux to study to what extent cutting-edge models can identify vulnerabilities and to demonstrate their security impact in a codebase of such high scale and importance.'

The Linux kernel supports cloud servers, Android devices, as well as many critical computing infrastructures around the world. The vulnerabilities identified in this OpenAI experiment are local privilege escalation vulnerabilities: attackers first need to have limited access to the machines with these vulnerabilities, but they may subsequently gain root privileges and gain full control over the systems.

948 vulnerabilities, 1169 exploitable codes

The agent of OpenAI scanned approximately 2.8 million lines of kernel code, which accounts for about 7% of the project. OpenAI states: "Our prototype identified and analyzed 2,308 candidate Linux kernels, dynamically reproduced 1,724 vulnerabilities, and generated 948 local privilege escalation (LPE) exploit codes. These codes may allow unauthorized local users to obtain root access rights."

After grouping the code based on the root causes and proposed fix solutions, researchers identified 948 different vulnerabilities.

This October's operation cost $1,213,430 in API credits, which averages out to about $1,279 per vulnerability. OpenAI compares this figure to the rewards of $20,000 to $70,000 in previous hacker competitions for privilege escalation vulnerabilities, stating that if the same results could be achieved manually, the theoretical value could range from $18.9 million to $66.3 million.

OpenAI also indicates that as early as July, earlier versions of the GPT-5.6 Sol solutions had identified 166 unique privilege escalation vulnerabilities at a cost of $513,690. In October, despite several months of patching for the kernel, an updated but yet-to-be-released cybersecurity model still found more than five times that number of vulnerabilities.

OpenAI also indicates that its agent utilized at least 10 technologies that are not familiar to researchers or external experts it consulted, rather than merely existing techniques retrieved from training data.

What if hackers also obtained the same tools?

AI The discovery of a critical vulnerability by the company is just another example that these systems are capable of conducting advanced network security research, a type of result that in the past usually required a great deal of human expertise.

Last week, Anthropic revealed that an unreleased version of Claude had discovered and exploited serious vulnerabilities in widely used software, including Chrome and Linux. Some of these vulnerabilities allowed for the remote execution of malicious code with almost no user interaction required.

Before the publication of OpenAI, the economic viability of such research was already undergoing changes.

Researchers from Vidoc Security used publicly available models to reproduce several vulnerability discoveries related to the powerful Mythos AI of Anthropic at a cost of less than $30 per file.

These tests indicate that it is becoming increasingly easier to identify vulnerabilities, but reproducing a vulnerability is not the same as developing reliable exploit code that can crack protected systems.

The latest experiments with OpenAI further automate this process, which could give defenders an advantage in patching software before criminals take advantage of the same vulnerabilities.

The company warns that attackers with limited access rights on outdated operating systems may increasingly be able to escalate their privileges by leveraging AI.

OpenAI is attempting to restrict access to its more robust network security model through verified security projects. Its Daybreak Red plan provides approved defenders with specialized AI tools for vulnerability research, exploitation testing, and mitigation solution development.