OpenAI Says Its Upcoming Cyber Model Found and Exploited 948 Linux Kernel Flaws at $1,279 Each

OpenAI on November 4 published a security report, A Thousand Fewer Paths to Root, describing an autonomous pipeline that finds vulnerabilities in the Linux kernel, proves they are exploitable, writes working exploits, and drafts patches for human review. The pipeline was driven by an upcoming cyber model the company has not yet released.

In an October run, analyst agents audited 5,034 targets spanning roughly 2.8 million lines of kernel code — about 7 percent of the codebase — and produced 2,308 candidate findings in 12 hours. Dynamic reproduction confirmed 1,724 of them, and exploit-development stages produced 291 information-leak proofs of concept plus 1,169 usable local privilege escalation exploits, which deduplicated to 948 unique LPE bugs.

The company put the total API cost at $1,213,430, with $775,929 going to exploit development and $316,549 to candidate analysis. That works out to about $1,279 per unique privilege escalation exploit, or roughly $359 per successful exploit once failed attempts are excluded. Fewer than 3.1 percent of unreproduced findings turned out to be analysis errors, according to the report.

The generational jump is the headline. In late July, OpenAI said, GPT-5.6 Sol found and exploited 166 unique LPEs for $513,690; the new model found more than five times as many novel LPEs despite three months of upstream kernel hardening in between. In orchestration experiments, GPT-5.5 Cyber needed a library of 103 exploitation techniques to lift its success rate from 76.2 percent to 88.9 percent, while GPT-5.6 Cyber gained only 1.5 points from the same assistance — meaning it can identify and apply techniques on its own. A sample of 166 exploits contained at least 10 novel exploitation techniques. All capability figures are company-reported and have not been independently verified.

In Google's hardened kernelCTF environment, the same approach produced 21 LPE exploits and 25 information-leak PoCs in 112 hours for about $32,936 — OpenAI notes that if all 21 were accepted, the theoretical bounty of roughly $1.5 million would be about 45 times the cost. Against desktop threat models used by Pwn2Own ($20,000) and TyphoonPWN ($70,000), it estimates the 948 exploits would be worth between $18.9 million and $66.3 million on the gray market, 16 to 55 times what the pipeline cost to build.

Disclosure is already moving. Within ten days of the scan and human triage, 61 of the LPEs were fixed upstream in Linux, 43 had staged patches, and 299 were in public discussion or reporting. Early runs also produced instructive failures: 20 of an initial 200 "successful" exploits were actually using other, unknown 0-day bugs to get root, and the kernelCTF run found and exploited a QEMU TCG bug without access to QEMU source code.

The report's practical warning is blunt: for a well-resourced attacker, access to a low-privileged account on a machine that is behind on patches should be treated as functionally equivalent to root. OpenAI argues that sandboxing, attack-surface reduction and formal verification will matter more as frontier models make vulnerability discovery cheap, and says it will keep expanding audits of critical open-source software through its Patch the Planet program.