OpenAI revealed that an autonomous pipeline of artificial intelligence agents identified and exploited 948 local privilege escalation vulnerabilities in the Linux kernel. Detailed in the company's official blog on November 4, 2026, the work cost roughly $1.2 million in API credits and demonstrates how frontier models can accelerate the discovery of flaws in critical software.
What happened?
The pipeline combines agents that scan code, reproduce flaws in isolated environments, generate proofs of concept, and propose patches. In a mid-October run, the system analyzed 2,308 candidates, dynamically reproduced 1,724 vulnerabilities, and produced 948 working local privilege escalation (LPE) exploits. These exploits allow a user with limited access to gain root privileges on a standard Linux distribution.
Additionally, 291 proofs of concept for information leaks were generated. The total cost was $1,213,430, or about $1,279 per validated LPE exploit. By comparison, an earlier July run with GPT-5.6 Sol found 166 vulnerabilities.
Why it matters
The Linux kernel underpins Android devices, cloud infrastructure, and supercomputers used by billions of people. With more than 40 million lines of code, it presents a broad attack surface. OpenAI notes that automated discovery scales faster than human triage and remediation capacity, making automated patch generation essential.
Linux maintainers had already patched dozens of the reported flaws during the research. The experiment took place in a controlled environment, and proposed patches require human review. The company is working with the Patch the Planet initiative to report and help fix the vulnerabilities.
What to expect
OpenAI argues that, in the near future, having a low-privileged account on an outdated system may effectively equate to root access for well-resourced actors. Defense in depth, such as sandboxes, becomes even more important. The work reinforces the need for high-assurance software and models trained to write secure code by default.
Source: OpenAI – A thousand fewer paths to root.
By Grok
