OpenAI cyber model produces 948 working Linux root exploits in one run
OpenAI published its write-up on November 4, 2026, describing an autonomous pipeline built on an unreleased cyber model. The run took place in mid-October against the latest Linux kernel, which contains more than 40 million lines of code. Scout agents narrowed the audit to 5,034 targets covering about 2.8 million lines, roughly 7 percent of the kernel. Analyst agents then produced 2,308 deduplicated candidate findings within 12 hours.
Triage agents tried to reproduce each candidate inside virtual machines running a kernel with KASAN enabled, and 1,724 were reproduced within 30 hours. Exploit agents produced 1,169 working root exploits and 291 proofs of concept for information leaks. Variant analysis grouped the root exploits by shared cause, which left 948 distinct vulnerabilities, each paired with a candidate patch. Tests ran on a stock desktop Linux setup with SMEP, SMAP and KASLR enabled. Total API spend came to $1,213,430, or about $1,279 per flaw.
According to AI Weekly, ten days after the scan, 61 flaws had been patched upstream, 43 had staging patches and 299 were under public discussion. OpenAI also said its agents used at least ten exploitation techniques it considered new. The write-up includes a caveat: in an early run, 20 of 200 claimed root exploits had reached root through a different zero-day than the one assigned.
Why it matters
For builders, any stock kernel or shared dependency should now be treated as exposed, since anyone able to rent a comparable agent pipeline could run the same search. The cost of about $1,279 per flaw sits far below the $20,000 to $70,000 that Pwn2Own and TyphoonPWN have historically paid for a single Linux escalation exploit. OpenAI estimates the 948 exploits would be worth $18.9 million to $66.3 million at those rates, or 16 to 55 times the run's cost.
For operators and investors, the bottleneck shifts from finding bugs to shipping fixes. Teams that run infrastructure should plan faster kernel patch cycles, keep inventories of shared dependencies, and verify exploit mitigations instead of assuming they hold. Companies selling AI security tooling or vulnerability research should expect buyers to compare their prices with API spend, and investors should ask how much of a security vendor's value depends on scarce human research. The 948 figure describes patches proposed by the pipeline, not fixes already shipped to Linux.
Key points
- 1The pipeline scanned about 2.8 million lines, roughly 7% of a kernel with more than 40 million lines.
- 2Analyst agents produced 2,308 candidate findings in 12 hours, and 1,724 were reproduced within 30 hours.
- 3Exploit agents produced 1,169 working root exploits, which grouped into 948 distinct vulnerabilities.
- 4Total API spend was $1,213,430, or about $1,279 per distinct flaw, according to the write-up.
