OpenAI says its autonomous security agents uncovered 948 distinct vulnerabilities in the Linux kernel, developed working exploits, and drafted fixes for human review—all for roughly $1.2 million in API credits, according to the company.

The October campaign cost $1,213,430 in credits, an average of about $1,279 per vulnerability, with a more capable, unreleased AI model ultimately finding more than five times as many flaws as an earlier version. Although the system also generated candidate patches, human review and software maintainers remain responsible for validating and implementing fixes.

"We chose Linux to study how effectively frontier models could find vulnerabilities and demonstrate their security impact in a codebase of this scale and importance," OpenAI said in results published Wednesday.

The Linux kernel powers cloud servers, Android devices, and much of the world's critical computing infrastructure. The flaws identified in the experiment were local privilege-escalation vulnerabilities: an attacker would first need limited access to a vulnerable machine, but could then potentially obtain root privileges and take full control. On shared systems such as cloud servers, that escalation step is precisely what turns a limited foothold into control over everything the machine runs.

948 flaws, 1,169 working exploits

OpenAI's agents scanned roughly 2.8 million lines of kernel code, about 7% of the project. "Our prototype identified and analyzed 2,308 candidate Linux kernel findings, dynamically reproduced 1,724 vulnerabilities, and produced 948 local privilege escalation (LPE) exploits that could allow an unprivileged local user to gain root access," OpenAI said.

After grouping the exploits by their underlying cause and proposed fix, researchers counted 948 distinct flaws, with exploits sharing the same root cause and remedy treated as one.

OpenAI compared the run's cost with historical payouts of $20,000 to $70,000 for Linux privilege-escalation exploits at hacking competitions, suggesting a hypothetical value of $18.9 million to $66.3 million if the same results had been found manually.

An earlier campaign by GPT-5.6 Sol found 166 unique privilege-escalation vulnerabilities in July for $513,690. The newer, unreleased cyber model found more than five times as many in October, even after the kernel had undergone months of fixes. OpenAI also said its agents used at least 10 exploitation techniques unfamiliar to the researchers and the outside experts they consulted—not merely existing tricks retrieved from training data.

What happens when hackers get the same tools?

The results add to growing evidence that AI systems are capable of performing advanced cybersecurity research of a kind that previously required significant human expertise.

Last week, Anthropic disclosed in a research report that an unreleased version of Claude had discovered and exploited serious vulnerabilities in widely used software, including Chrome and Linux. Some of those exploits could execute malicious code remotely with little or no user interaction. Published within roughly a week of each other, the two reports underscore how quickly frontier labs are moving to publicize the security capabilities of their most capable models while keeping those systems out of general release.

The economics of this kind of research were already shifting before OpenAI's announcement. Researchers at Vidoc Security reproduced several vulnerability discoveries associated with Anthropic's powerful Mythos AI using publicly accessible models for less than $30 per file, as Decrypt previously reported.

Those tests showed that finding vulnerabilities was becoming increasingly accessible, although reproducing a flaw is different from developing a reliable exploit capable of compromising a protected system.

OpenAI's latest experiment goes further by automating much of the exploitation process, a development that could give defenders an advantage in patching software before criminals are able to exploit the same weaknesses. The company warned that attackers with limited access to an outdated operating system may increasingly be able to escalate their privileges with AI assistance. For now, the candidate fixes drafted by the agents remain in human reviewers' hands—making the handling of AI-generated patches for widely used code a concrete next development to watch.

OpenAI is seeking to restrict access to its more capable cybersecurity models through verified security programs. Its Daybreak Red initiative provides approved defenders with access to specialized AI tools for vulnerability research, exploit testing, and mitigation development.