OpenAI says unreleased AI model exposed 948 cyber flaws in Linux, cost $1.2 million

AI models seem to be only getting better at cybersecurity. OpenAI has announced that an unreleased AI model managed to find 948 unique vulnerabilities in Linux. The AI then used these flaws to develop exploits that would give a user root access. This was part of OpenAI’s automated cybersecurity research pipeline that ran last month, and cost over $1.2 million in API credits.

OpenAI disclosed the findings in a post on X. “Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline,” the company said. As per OpenAI, the aim was to help defenders find and fix vulnerabilities before attackers could use them. Though OpenAI has not given any details of the upcoming model, including a possible release date.

Linux kernel is one of the most-used pieces of software in the world. It forms the basis of Android devices, cloud infrastructure, as well as supercomputers. OpenAI says billions of people use devices running on Linux globally, with the project having over 40 million lines of code.

How did OpenAI’s new model find flaws in Linux?

In a blog post, OpenAI said that the unreleased model managed to find flaws and build exploits in a single run in mid-October. As part of the run, the AI first identified code that was relevant to cybersecurity and then scanned it for potential flaws. Later, it reproduced those flaws in isolated environments – away from the internet – and then tested whether they can be exploited by a user.

OpenAI said the agents also discovered and used at least 10 Linux exploitation techniques that its researchers and consulted experts considered novel.

As per the company, “scout agents” identified 5,034 security-relevant targets covering about 2.8 million lines of Linux kernel code, or roughly 7 per cent of the codebase. Analyst agents examined those targets and produced 2,308 candidate findings within 12 hours. The system then attempted to reproduce each issue inside isolated virtual machines and confirmed 1,724 vulnerabilities within 30 hours.

Reproducing a vulnerability established that the flaw was real but did not necessarily show that it could be used for privilege escalation – where a user can gain root access. A screening agent narrowed the confirmed findings to 1,238 local privilege escalation candidates and 313 information-leak candidates.

Exploit-development agents tested the flaws under standard security protections. The system generated proofs of concept for 291 information leaks and 1,169 working local privilege escalation exploits.

Then, the AI agents found that the 1,169 exploits mapped to 948 distinct patches, which OpenAI used as its count of unique local privilege escalation vulnerabilities. The distinction is important because several exploits could target the same underlying flaw. Human security researchers reviewed the results before reporting vulnerabilities and proposing fixes to Linux maintainers.

Up to 55 times cheaper than humans

As per OpenAI, the full campaign cost $1,213,430 in API credits. The company compared that amount with historical Linux kernel bounty rates given to users, which are $20,000 at Pwn2Own and $70,000 at TyphoonPWN. Applying those rates to 948 vulnerabilities produced a hypothetical value of $18.9 million to $66.3 million, or about 16 to 55 times the pipeline's cost.

On an adapted kernelCTF environment with a reduced attack surface, the pipeline generated 25 information-leak proofs of concept and 21 local privilege escalation exploits. OpenAI said that only 26 confirmed zero-day local privilege escalation exploits had been accepted by kernelCTF in 2026 so far.

That campaign cost about $32,936 in API credits. OpenAI estimated that the 21 vulnerabilities could have resulted in about $1.5 million in bounty awards if accepted by the programme, or roughly 45 times the scanning cost.

The comparison represents an illustrative valuation rather than actual bounty income. Though it appears that AI agents can be drastically cheaper in finding flaws when compared to human experts, at least from this data.

OpenAI stated that it sent hundreds of reports and patches for critical vulnerabilities in line with guidance from the Linux kernel community. It is also working with Linux maintainers and partners under its Patch the Planet initiative to report and address additional flaws.

OpenAI said the finding suggested that frontier large language models could discover and patch new approaches to exploitation rather than merely apply known techniques. It also noted that independent checks remained necessary. In one early run, 20 of 200 apparently successful exploits had gained root access through vulnerabilities other than the ones assigned to them.

The October campaign substantially exceeded the results of an earlier run. In July, GPT-5.6 Sol found and exploited 166 unique local privilege escalation vulnerabilities at a cost of $513,690.

Previously, OpenAI’s models also helped find flaws in the programming language PHP that put the internet on alert. Meanwhile, Anthropic’s Claude Mythos has also found thousands of critical flaws across various systems since its release in April this year.

- Ends