OpenAI AI agents discovered 948 ways to gain root access in Linux
According to the company, the autonomous system identified 948 unique privilege escalation vulnerabilities and prepared draft fixes. The review cost approximately $1.2 million in API credits
OpenAI presented the results of an experiment on the automated search for vulnerabilities in the Linux kernel. The system, based on a cyber model, combines agents that select code sections, search for errors, reproduce them in an isolated environment, and verify the possibility of exploitation.
According to the company, in the October launch, the system analyzed about 2.8 million lines—approximately 7% of the kernel's codebase. Out of 2,308 potential issues, 1,724 were successfully reproduced.
For 948 unique vulnerabilities, the agents created working methods for local privilege escalation. Such errors could allow a user with limited permissions to gain root access. This does not imply the possibility of remote hacking of any Linux computer.
The system also prepared draft patches. These are reviewed by researchers and passed on to kernel developers: the presence of a proposed fix does not yet mean it has been accepted and delivered to users.
The total cost of the launch amounted to $1,213,430 in API credits—approximately $1,279 per unique privilege escalation vulnerability.
OpenAI also notes the limitations of automation: in an earlier launch, some exploits used different errors instead of the targeted ones. After manual verification, an independent exploitation path check was added to the system.
