OpenAI pipeline produced 948 Linux root exploits in 30 hours

OpenAI has described an automated pipeline that used an upcoming cyber-focused artificial intelligence model to find, reproduce and exploit local privilege escalation flaws in the Linux kernel, according to a report published by The Hacker News on 5 November 2026. The company detailed the work on Wednesday 4 November 2026. In a single 30-hour run, the pipeline produced 948 local privilege escalation exploits at a cost of $1.2 million in API usage.

Local privilege escalation means an attacker who already has limited access to a machine can raise that access to root, the highest level of control on a Linux system. It is one of the most common steps in turning a small foothold into full compromise of a server or workstation.

According to the report, the prototype identified and analysed 2,308 candidate Linux kernel findings. From that pool it dynamically reproduced 1,724 vulnerabilities, meaning it ran them and confirmed the faulty behaviour rather than simply flagging suspicious code. It then produced 948 working local privilege escalation exploits. OpenAI said the pipeline was used to validate and patch the vulnerabilities it handled.

The numbers are notable mainly for their ratio. Roughly three quarters of the candidate findings were confirmed by running them, and more than half of the original candidates ended in a working exploit. Static analysis tools have long generated large volumes of candidate issues that human engineers must triage by hand. The reported figures suggest the pipeline carried out much of that triage itself, including the reproduction step that usually takes the most engineering time.

The cost figure is also unusual for a security research disclosure. At $1.2 million in API costs for the run, the pipeline worked out at roughly $1,266 per exploit produced, based on the figures OpenAI gave. That is a price point the company chose to publish rather than one derived by outside analysts, and it covers API usage only. The report does not break down what other infrastructure or staff time the run required.

The model used is described as upcoming and cyber-focused, so it is not generally available. OpenAI has not, in the material reported, said when or whether the model will be released, or on what terms access would be granted.

For defenders, the practical detail is the patching claim. OpenAI said the pipeline both validated and patched the vulnerabilities, which places the work in the same category as other automated bug-finding and bug-fixing efforts aimed at large open source codebases. The report does not state how many of the 948 exploits correspond to previously unknown flaws, nor how the fixes were handled with Linux kernel maintainers. Those details matter for judging the practical effect of the run, and they are not covered in the available account.