OpenAI stated that its autonomous security agents discovered 948 distinct vulnerabilities in the Linux kernel, developed functional exploit code, and drafted remediation proposals for review—all within approximately $1.2 million worth of API credits.
OpenAI stated in its results released on Wednesday: "We chose Linux to study how far frontier models can discover vulnerabilities and to demonstrate their security implications in a codebase of such scale and importance."
The Linux kernel powers cloud servers, Android devices, and many critical computing infrastructures worldwide. The vulnerability identified in OpenAI’s experiment is a local privilege escalation vulnerability: an attacker must first have limited access to the vulnerable machine, but could then gain root privileges and fully control the system.
948 vulnerabilities, 1,169 available exploit codes
OpenAI's agent scanned approximately 2.8 million lines of kernel code, accounting for about 7% of the project. OpenAI stated: “Our prototype identified and analyzed 2,308 Linux kernel candidate findings, dynamically reproduced 1,724 vulnerabilities, and generated 948 local privilege escalation (LPE) exploits that could allow unauthorized local users to gain root access.”
After grouping the exploit codes by root cause and proposed fix, researchers identified 948 distinct vulnerabilities.
This October’s campaign consumed $1,213,430 in API credits, equivalent to approximately $1,279 per vulnerability. OpenAI compared this figure to historical bounties for Linux privilege escalation vulnerabilities in hacker competitions, which range from $20,000 to $70,000, noting that the theoretical value of finding the same results manually could range between $18.9 million and $66.3 million.
OpenAI also stated that as early as July, an earlier version of the GPT-5.6 Sol solution identified 166 unique privilege escalation vulnerabilities at a cost of $513,690. In October, despite the kernel having undergone months of patches, a newer, yet unreleased, cybersecurity model found more than five times as many vulnerabilities.
OpenAI also stated that its agents used at least ten exploitation techniques unfamiliar to researchers and external consultants, rather than merely retrieving existing techniques from training data.
What if hackers also get the same tools?
An AI company discovered a critical vulnerability, further demonstrating that these systems can perform advanced cybersecurity research—results that previously required significant manual expertise.
Last week, Anthropic disclosed that one of its unreleased Claude versions discovered and exploited critical vulnerabilities in widely used software, including Chrome and Linux. Some of these vulnerabilities allowed remote execution of malicious code with minimal user interaction.
The economics of this type of research had already been changing before OpenAI's announcement.
Researchers at Vidoc Security used publicly available models to reproduce several vulnerability discoveries related to Anthropic's powerful Mythos AI, at a cost of less than $30 per file.
These tests show that discovering vulnerabilities is becoming increasingly accessible, but reproducing a vulnerability is not the same as developing reliable exploit code capable of breaking into protected systems.
OpenAI's latest experiment further automates this exploitation process, potentially giving defenders an advantage in patching software before criminals can exploit the same vulnerabilities.
The company warned that attackers with limited access on outdated operating systems may increasingly be able to use AI to escalate privileges.
OpenAI is attempting to restrict access to its more powerful cybersecurity models through verified security programs. Its Daybreak Red initiative provides approved defenders with specialized AI tools for vulnerability research, exploitation testing, and mitigation strategy development.