OpenAI Pipeline: 948 Linux Kernel LPEs at $1,279 Each

OpenAI's November 4, 2026 security report describes an autonomous Linux kernel vulnerability pipeline spanning target selection, dynamic validation, exploitability assessment, and patch generation. In a mid-October 2026 run on a commonly used Linux distribution, the pipeline produced 948 unique local privilege escalation findings backed by 948 distinct patches. OpenAI reports the run cost $1,213,430 in API credits, or $1,279 per LPE. Applying the historical Pwn2Own and TyphoonPWN Linux kernel LPE bounties of $20,000 and $70,000 yields a hypothetical valuation of $18.9M–$66.3M, about 16x to 55x the pipeline cost.

Pipeline and cost distribution

The pipeline has five stages: vulnerability discovery, dynamic validation, exploitability assessment, patch generation, and human review. Scout agents selected 5,034 audit targets covering about 2.8 million lines, roughly 7% of the kernel, for $3,426. Analyst agents then produced 2,308 deduplicated candidate findings in 12 hours for $316,549. A KASAN-instrumented triage stage reproduced 1,724 vulnerabilities, 74.7% of candidates, in 30 hours for $73,330. Static screening narrowed the set to 313 information-leak and 1,238 LPE candidates. ExploitDev agents worked in VMs with SMEP, SMAP, KASLR, and HARDENED_USERCOPY enabled; that stage cost $775,929, including $691,000 for LPE candidates and $84,929 for information leaks. The agents produced 291 information-leak PoCs and 1,169 working LPE exploits; variant analysis and 948 distinct patches defined the final unique-LPE count.

On an adapted kernelCTF environment, a 112-hour run cost $32,936, with vulnerability discovery taking 83.7% of that spend. It produced 25 information-leak PoCs and 21 LPE exploits. OpenAI notes that 26 confirmed 0-day LPEs had been accepted by kernelCTF in all of 2026 at the time of writing. If accepted, the 21 would represent roughly $1.5M in bounty awards, about 45x the scan cost.

Model capability and orchestration

OpenAI reports that in late July 2026, GPT‑5.6 Sol found and exploited 166 unique LPEs for $513,690. The mid-October run with an unnamed upcoming cyber model produced 948 for $1,213,430—more than five times as many for about 2.4 times the spend. In a separate orchestration experiment, OpenAI compared GPT‑5.5 Cyber and GPT‑5.6 Cyber. Giving GPT‑5.5 Cyber a formatted library of 103 exploitation techniques raised its success rate on 63 candidates from 76.2% to 88.9%. The same technique layer added only 1.5 percentage points for GPT‑5.6 Cyber, and the lightly orchestrated GPT‑5.6 workflow outperformed the heavily scaffolded GPT‑5.5 workflow.

Operational findings

OpenAI reports that when the scan and manual analysis finished after 10 days, 61 LPEs had already been patched upstream, 43 had staging patches, and 299 were being publicly discussed, so automated triage must revalidate against the latest tree. In an early run, 20 of 200 reported exploitation successes achieved root through a different zero-day than the assigned vulnerability; OpenAI added an independent check confirming each exploit used its designated bug. During kernelCTF work, an analyst agent found a QEMU TCG privilege escalation without QEMU source access; it was fixed upstream but not in the release build used. Patch quality improved through iteration from functionally correct but overlong or incomplete fixes to minimal patches consistent with kernel conventions; in a few cases the pipeline caught errors in human-expert patches before merge.

AI Mastery analysis

The most operationally significant number is $1,279 per LPE on the stock kernel, alongside the kernelCTF campaign cost of about $1,568 per LPE. That converts "AI can find bugs" into a procurement question: at what per-finding cost does automated re-scanning become mandatory before shipping kernel updates or container runtimes? The gap to the $20,000 Pwn2Own floor means the same economics are available to well-resourced threat actors, not only defenders. This matches our agent security analysis: deployed software surfaces are now subject to systematic, affordable enumeration.

The orchestration result also matters. Scaffolding worth 12.7 points for GPT‑5.5 Cyber added 1.5 points for GPT‑5.6 Cyber. Harness-heavy security pipelines accumulate orchestration debt that stronger base models can make redundant, inverting the usual pipeline-architecture claim that harness design drives most gains. OpenAI says the reported costs are upper bounds because the work prioritized methodology and model capability over cost optimization, so the per-LPE number has room to fall.

The evidence is heavier than a benchmark claim: a named codebase, a Pwn2Own-style threat model, per-stage dollar totals, 10 novel techniques sampled from 166 exploits and shared with third-party kernel experts, and patch reviews. That makes the release a reproducibility target for independent researchers rather than only a vendor capability assertion.

Sources

• Primary: OpenAI News: A thousand fewer paths to root