OpenAI revealed on Wednesday (4) that an autonomous artificial intelligence system managed to develop 948 distinct exploits capable of elevating privileges to root level in the Linux kernel. The experiment, conducted in October with an as-yet-unreleased cybersecurity model, cost US$ 1.21 million in API credits and also produced fix proposals for the identified vulnerabilities.
The results were presented in the technical report A Thousand Fewer Paths to Root. According to the company, the system analyzed 2,308 possible security flaws, reproduced 1,724 vulnerabilities in isolated environments, and generated 291 proofs of concept for information leakage.
In the exploitation stage, the agents produced 1,169 functional exploits. After an analysis of variants and necessary fixes, OpenAI identified 948 distinct local privilege-escalation vulnerabilities.
These flaws allow someone with limited access to a system to obtain administrator permissions. The tests were run on virtual machines with a standard Linux distribution and security protections enabled, without deliberately weakening the environment.
The average cost of the complete process was US$ 1,279 per exploitable vulnerability, considering all stages of discovery, validation, exploitation, and fix generation.
Model found more than five times the flaws identified in July
The difference compared with previous experiments is significant. In July 2026, a run with GPT-5.6 Sol had identified and exploited 166 distinct privilege-escalation vulnerabilities, at a cost of US$ 513,690.
In the October run, the latest model found more than five times that number, even after three additional months of fixes and security improvements in the Linux kernel.
OpenAI also examined a sample of 166 exploits and identified at least ten exploitation techniques that, according to the company, were unknown even to the outside experts consulted. The technical details of these approaches were not disclosed for security reasons.
Automation did not eliminate human participation. The system was designed to discover flaws, prove their exploitation, and propose patches, but researchers remain responsible for reviewing the results before forwarding them to the Linux maintainers.
Hundreds of fixes still depend on review
The discovery of the vulnerabilities does not mean that all of them have been fixed. OpenAI said it had forwarded hundreds of reports and patch proposals, prioritizing the most critical problems and working with partners in the initiative Patch the Planet to address the rest.
At the end of the ten days of execution and analysis, 61 vulnerabilities had already received fixes incorporated into the kernel, another 43 had patches in preparation, and 299 were being discussed publicly. These numbers also reflect the independent work of the Linux developer community.
The experiment shows the stated ability of a specialized model to automate complex security research tasks, but its results remain tied to the configurations tested and the data disclosed by OpenAI itself.
The company advocates combining automated discovery, patch generation, and human review to reduce the interval between identifying a vulnerability and fixing it.
