OpenAI has revealed that an unreleased artificial intelligence model discovered 948 unique vulnerabilities in the Linux kernel and developed exploits capable of granting users root access. The findings emerged from an automated cybersecurity research campaign conducted last month, which reportedly cost more than $1.2 million in API credits.
The company shared the results on X, stating that its security researchers had developed an autonomous pipeline for discovering, exploiting and patching cybersecurity vulnerabilities. The initiative was designed to help security teams identify and address weaknesses before malicious actors could exploit them.
OpenAI has not disclosed the identity of the unreleased model or provided a potential launch timeline.
The Linux kernel is among the world's most widely used software components. It powers Android devices, cloud infrastructure and supercomputers. According to OpenAI, billions of people rely on Linux-based devices, while the project itself contains more than 40 million lines of code.
How did OpenAI's unreleased AI model identify Linux vulnerabilities?
In a blog post, OpenAI explained that the model discovered and developed exploits for vulnerabilities during a single research run in mid-October.
The AI first identified cybersecurity-relevant sections of the codebase and searched for potential weaknesses. It subsequently reproduced the suspected vulnerabilities in isolated environments without internet access and tested whether they could be exploited.
OpenAI said its agents also identified and used at least 10 exploitation techniques that researchers and external experts consulted by the company considered novel.
The research pipeline followed several stages:
Scout agents identified 5,034 security-relevant targets covering approximately 2.8 million lines of Linux kernel code, representing around 7 per cent of the codebase.
Analyst agents examined the targets and generated 2,308 potential findings within 12 hours.
The system attempted to reproduce the issues inside isolated virtual machines, confirming 1,724 vulnerabilities within 30 hours.
Reproducing a vulnerability established that a flaw existed, but it did not necessarily demonstrate that the issue could be exploited to gain root access.
A screening agent subsequently reduced the confirmed findings to 1,238 potential local privilege escalation vulnerabilities and 313 potential information leaks.
Exploit-development agents then tested the vulnerabilities under standard security protections. The system generated proofs of concept for 291 information leaks and 1,169 working local privilege escalation exploits.
The 1,169 exploits were ultimately linked to 948 distinct patches. OpenAI used this figure to determine the number of unique local privilege escalation vulnerabilities discovered, since multiple exploits could target the same underlying flaw.
Human security researchers reviewed the findings before the vulnerabilities were reported and fixes were proposed to Linux maintainers.
AI cybersecurity research could be up to 55 times cheaper than humans
OpenAI said the entire research campaign consumed $1,213,430 in API credits.
The company compared this expenditure with historical Linux kernel vulnerability bounty rates, including $20,000 at Pwn2Own and $70,000 at TyphoonPWN. Applying these rates to the 948 vulnerabilities resulted in a hypothetical valuation of $18.9 million to $66.3 million.
According to OpenAI, this translates into a potential value between 16 and 55 times the cost of the automated research pipeline.
The company also conducted a separate campaign in an adapted kernelCTF environment with a smaller attack surface. The system generated 25 information-leak proofs of concept and 21 local privilege escalation exploits.
OpenAI noted that kernelCTF had accepted only 26 confirmed zero-day local privilege escalation exploits in 2026 at the time of its research.
The smaller campaign cost approximately $32,936 in API credits. OpenAI estimated that the 21 vulnerabilities could potentially have earned around $1.5 million in bounty rewards if accepted, representing approximately 45 times the research expenditure.
However, these figures are illustrative estimates rather than actual earnings. The vulnerabilities would still need to meet the relevant bounty programme's requirements and receive acceptance.
OpenAI is working with Linux maintainers to patch vulnerabilities
OpenAI said it had submitted hundreds of reports and patches for critical vulnerabilities in accordance with Linux kernel community guidelines.
The company is also collaborating with Linux maintainers and partners through its Patch the Planet initiative to identify and resolve additional security weaknesses.
OpenAI believes the findings demonstrate that advanced AI models can discover new exploitation techniques and develop patches rather than simply relying on previously known vulnerabilities.
However, the company acknowledged that independent verification remains essential. During an early research run, 20 out of 200 apparently successful exploits gained root access by exploiting vulnerabilities different from those originally assigned to them.
The October campaign also significantly outperformed an earlier experiment. In July, GPT-5.6 Sol reportedly discovered and exploited 166 unique local privilege escalation vulnerabilities at a cost of $513,690.
AI models are increasingly finding critical software vulnerabilities
OpenAI's Linux research follows other cybersecurity findings involving AI models.
The company's models previously helped identify vulnerabilities in PHP, a programming language widely used across the internet.
Meanwhile, Anthropic's Claude Mythos reportedly discovered thousands of critical vulnerabilities across different systems following its release in April.
The latest findings suggest that AI-powered cybersecurity research is advancing rapidly. While automated systems could help developers identify and patch vulnerabilities more efficiently, their ability to generate working exploits also raises concerns about how such capabilities could be misused by malicious actors.
