OpenAI reported that a vulnerability research pipeline, driven by an upcoming cyber model, dynamically reproduced 1,724 Linux kernel vulnerabilities and generated 948 distinct local privilege escalation (LPE) exploits during a mid‑October 2026 run. The project bundled proofs of concept, root cause analyses and candidate patches for maintainer review.
Pipeline and findings
The prototype operates in five stages: vulnerability discovery, dynamic validation, exploitability assessment, patch generation and human review. Scout agents first selected 5,034 audit targets—about 2.8 million lines of code, roughly 7% of the kernel—at a reported cost of $3,426 in API credits. Analyst agents then produced 2,308 candidate findings in about 12 hours, a stage that cost $316,549.
Triage agents attempted to reproduce each candidate in isolated virtual machines with KASAN enabled. The system reproduced 1,724 vulnerabilities, a ≈74.7% reproduction rate, at a validation cost of $73,330; OpenAI reported that 3.1% of candidates reflected incorrect analysis.
An initial screening reduced the set to 313 information‑leak candidates and 1,238 LPE candidates. Exploit‑development agents produced 291 information‑leak PoCs and 1,169 working LPE exploits in VMs configured with SMEP, SMAP, KASLR and HARDENED_USERCOPY. Variant analysis yielded 948 distinct patches corresponding to unique LPEs.
Costs, comparisons and observations
The full run consumed $1,213,430 in API credits, about $1,279 per unique LPE. OpenAI reported successful exploit development averaged $359 in credits; a small set of unsuccessful, hard cases averaged much higher costs. On an adapted kernelCTF environment the pipeline obtained 25 information‑leak PoCs and 21 LPE exploits for roughly $32,936 in API credits.
OpenAI noted that earlier work with GPT‑5.5 Cyber had produced 24 LPE exploits and that the newer cyber model discovered more than five times as many novel LPEs as a GPT‑5.6 Sol run in July, despite several months of hardening. The agents reportedly discovered at least 10 novel Linux exploitation techniques. OpenAI added that maintainers patched over 100 LPEs within 10 days while about 300 more were publicly discussed, and that the team reported findings alongside candidate patches and worked with Patch the Planet partners.
OpenAI characterized defense in depth (for example, sandboxes such as Firecracker and gVisor) and automated patch generation as important elements to limit exploitation windows. The campaign focused on measuring model capability and methodology rather than cost optimization.
Original source: OpenAI News