OpenAI has published the results of a striking security experiment: an autonomous pipeline powered by an upcoming cyber model scanned the Linux kernel, proved hundreds of vulnerabilities were real, built working exploits for them, and drafted a patch for each one. In a single mid-October run, the system produced 948 distinct local privilege escalation vulnerabilities, the kind of flaw that lets an ordinary logged-in user gain root access, the highest level of control on a Linux machine.
The numbers behind the run show how the work was narrowed down. Scout agents first identified 5,034 audit targets covering about 2.8 million lines of code, roughly 7 percent of the kernel. Analyst agents then produced 2,308 candidate findings within 12 hours. Triage agents reproduced 1,724 of those vulnerabilities in isolated virtual machines, about 74.7 percent of the candidates, and only 3.1 percent turned out to rest on an incorrect analysis. Exploit-development agents went further, producing 1,169 working privilege escalation exploits and 291 information-leak proofs of concept, which variant analysis then consolidated into the 948 distinct vulnerabilities, each with its own patch.
OpenAI was unusually open about cost. The full run consumed 1,213,430 dollars in API credits, about 1,279 dollars per distinct vulnerability, and a successful exploit cost an average of 359 dollars to develop. The company noted that those figures are closer to an upper bound than a floor, because better harness design could reduce them. For comparison, it valued the same exploits at 18.9 million to 66.3 million dollars using historical bug bounty amounts from competitions such as Pwn2Own and TyphoonPWN.
The result also shows how quickly model capability is moving. In June, GPT-5.5 Cyber had produced 24 privilege escalation exploits under OpenAI's Patch the Planet initiative. A late July run with GPT-5.6 Sol found 166 unique vulnerabilities. The mid-October run, three months later and against a kernel that had been hardened in the meantime, found 948. In a sample of 166 exploits, the agents also used 10 exploitation techniques that outside kernel experts said were new to them, a detail OpenAI withheld in full for safety reasons.
None of this bypassed human judgment. Researchers reviewed each exploit, patch, and proposed severity before anything was reported to kernel maintainers, and that review caught a real problem: in an early run, 20 supposedly successful exploits had actually reached root through a different vulnerability than the one assigned, which led the team to add an independent check that each exploit used its assigned flaw. By the time the scan and manual analysis finished 10 days later, 61 of the flaws had already been patched upstream, 43 had staging patches, and 299 were being reported and discussed publicly.
The practical lesson is less about one model than about pace. Linux underpins Android devices, cloud infrastructure, and supercomputers used by billions of people, and OpenAI argues that defenders should now assume an unpatched system with a low-privileged account is close to root access for a well-resourced attacker. Keeping systems updated, and rescanning familiar code as models improve, just became a much stronger recommendation.