# OpenAI pipeline finds 948 Linux privilege escalation vulnerabilities at $1.2 million cost > OpenAI's autonomous pipeline found 948 unique Linux kernel LPE exploits at an API cost of $1.2 million, averaging $1,279 per working exploit. - Type: AI news - Page: https://completeaitraining.com/news/openai-pipeline-finds-948-linux-privilege-escalation/ - Published: 2026-11-05 - Updated: 2026-11-05 - Author: Jeroen Erne - Jobs: Healthcare, Insurance, IT and development, Operations, Science and research - Reading time: 3 min An autonomous vulnerability discovery pipeline built by OpenAI found and validated 948 unique local privilege escalation (LPE) exploits in the Linux kernel, using an upcoming cyber model at an API cost of $1.2 million. The run, completed in mid-October 2026, demonstrates how security research that once required separate, sustained investigations for each finding can now be run as a systematic, automated process. The pipeline identified 2,308 candidate kernel findings, dynamically reproduced 1,724 vulnerabilities, and produced working exploits that could allow an unprivileged local user to gain root access on a standard Linux desktop distribution. The total cost worked out to $1,279 per LPE exploit. Applying historical bug bounty rates from programs like Pwn2Own ($20,000) and TyphoonPWN ($70,000) to the 948 exploits yields a hypothetical valuation of $18.9 million to $66.3 million, roughly 16 to 55 times the pipeline's cost. #### How the five-stage pipeline works The research prototype operates across five stages. First, Scout agents identify high-value audit targets by reading through kernel folders and files, flagging code reachable by low-privileged users. This narrowed the audit to roughly 7% of the kernel's 40 million lines of code. Analyst agents then scan those targets for candidate vulnerabilities, producing 2,308 findings after deduplication within 12 hours at a cost of $316,549 in API credits. Next, triage agents attempt to reproduce each candidate finding inside an isolated virtual machine running a kernel with Kernel Address Sanitizer (KASAN). Within 30 hours, the system reproduced 1,724 vulnerabilities, about 74.7% of the candidates. Only 3.1% of the remaining findings reflected incorrect model analysis. The validation stage cost $73,330. The exploitability assessment stage uses a screening agent to filter validated vulnerabilities by security impact, then assigns ExploitDev agents to build full working exploits. These agents operated against a stock desktop operating system with standard protections enabled, including SMEP, SMAP, KASLR, and HARDENED\_USERCOPY. The exploitation stage cost $775,929 in total, averaging $359 per successful exploit. Finally, the system generates candidate patches and a human security researcher reviews the outputs before reporting to maintainers. #### Maintainers are patching faster than the scan runs The scan ran against the latest Linux kernel in mid-October 2026. Within the 10 days it took to complete the autonomous scan and manual review, 61 of the LPEs had already been patched upstream, 43 had staging patches, and 299 were being publicly discussed. The rapid community response underscores why automated triage must revalidate findings against the latest code to avoid duplicate or obsolete reports. "Within the next few years, we believe that if there are vulnerabilities in a given software project, frontier models will likely be able to discover them and even less capable models may be able to develop exploits once disclosed," the team wrote. "In a world of no unknown vulnerabilities, it will be critical to build and maintain high assurance software." #### Stronger models find more vulnerabilities with less orchestration An earlier July 2026 run using GPT-5.6 Sol found and exploited 166 unique LPEs at a cost of $513,690. The mid-October run with the newer cyber model found 948 LPEs, more than five times as many, despite three months of kernel hardening between scans. The team also found that stronger models require lighter orchestration. A heavily scaffolded workflow that improved GPT-5.5 Cyber's success rate from 76.2% to 88.9% across 63 exploitation candidates provided only a marginal 1.5 percentage-point improvement when applied to GPT-5.6 Cyber. The agents also demonstrated novel exploitation techniques. A sample of 166 exploits revealed 10 techniques previously unknown to third-party kernel exploitation experts, including new persistent elastic objects for cache spraying, new page grooming methods, and new data-only privilege escalation approaches. #### Why this matters for IT and development professionals For teams managing Linux infrastructure, the findings carry a direct operational message: a low-privileged account on an operating system that is not current on patches should be considered functionally equivalent to root access for a well-resourced threat actor. Defense in depth through sandboxing and isolation technologies like Firecracker and gVisor becomes essential for limiting blast radius. The cost data also suggests that defenders who proactively scan and patch their own software can make exploitation cost-prohibitive-the pipeline's kernelCTF run found 21 LPE exploits at an API cost of roughly $32,936, which would yield approximately $1.5 million in bounty awards if accepted. For professionals building security testing workflows, the research shows that investing in automated validation and patch generation pipelines is quickly becoming a practical necessity rather than a speculative research project. --- This is the Markdown form of https://completeaitraining.com/news/openai-pipeline-finds-948-linux-privilege-escalation/. The whole site is mapped at https://completeaitraining.com/llms.txt.