On November 4 OpenAI published "A thousand fewer paths to root." Its autonomous pipeline for vulnerability discovery, validation, exploitation and patching scanned the Linux kernel in mid-October with an upcoming cyber model, reproduced 1,724 vulnerabilities and produced exploits and patches for 948 distinct local privilege escalation (LPE) bugs. The run cost $1,213,430 in API credits, about $1,279 per LPE, and hundreds of reports and patches have already gone to maintainers.
Key Takeaways
- ✓Five-stage pipeline: discovery, dynamic validation, exploitability assessment, patch generation, human review
- ✓Of 2,308 candidate findings, 1,724 were reproduced, yielding 291 info-leak PoCs and 948 distinct LPEs
- ✓Total cost $1,213,430; historical Pwn2Own and TyphoonPWN bounties imply a hypothetical $18.9M to $66.3M value
- ✓The newer model found more than 5x the novel LPEs of a July GPT-5.6 Sol run (166)
- ✓The agents used at least 10 novel Linux exploitation techniques, withheld for safety reasons
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the Industry Pain Point
Frontier models make vulnerability discovery easier to scale, but validation and prioritization are the real bottleneck: thousands of candidate findings are far too many to review by hand. In June, under its Patch the Planet initiative, OpenAI used GPT-5.5 Cyber to build 24 Linux kernel local privilege escalation (LPE) exploits. It has now turned that into a full pipeline and published the results in an official post on November 4.
Architecture Highlights and Mechanism
The pipeline is built around an upcoming cyber model and has five stages. Scout agents pick security-relevant code reachable by low-privileged users, 5,034 audit targets covering about 2.8 million lines, roughly 7% of the kernel. Analyst agents scan them and produce 2,308 candidates. Triage agents reproduce findings in isolated VMs with KASAN. ExploitDev agents attempt exploits on a standard distribution configuration with SMEP, SMAP and KASLR enabled. Finally the system generates patches, and human security researchers review everything before reporting to maintainers.
Benchmarks and Measured Results
According to OpenAI, the mid-October run reproduced 1,724 vulnerabilities and produced 291 information-leak PoCs and 948 distinct LPEs for $1,213,430 in API credits, about $1,279 each. A July run of the same process with GPT-5.6 Sol found 166 LPEs for $513,690. On the hardened kernelCTF environment it produced 21 LPEs for about $32,936; for comparison, kernelCTF has accepted 26 confirmed 0-day LPEs so far in 2026.
Getting Started for Developers
Following the kernel community's guidance, OpenAI has already sent hundreds of reports and patches for critical bugs and is sending the rest through Patch the Planet partners; during the scan, maintainers had already patched 61 LPEs and staged patches for 43 more. OpenAI recommends keeping kernels up to date, treating a low-privileged account on an unpatched system as equivalent to root, using sandboxes such as Firecracker and gVisor for defense in depth, and rescanning old code as models improve. The cyber model itself has not been released yet.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.