OpenAI Says an Autonomous Pipeline Built 948 Linux Root Exploits, With Patches, for $1.2 Million in API Credits
OpenAI said on November 4 that an automated pipeline running an unreleased cyber model reproduced 1,724 Linux kernel vulnerabilities in a mid-October run and produced 948 distinct local privilege escalation exploits, each with a candidate patch, for $1,213,430 in API credits, about $1,279 per exploit. OpenAI says humans reviewed the output and it has sent hundreds of reports and patches to maintainers. All figures are OpenAI's own.
Finding a way to root on Linux is turning into a line item. In a research post on November 4, OpenAI said an automated pipeline running an upcoming, unreleased cyber model produced 948 distinct local privilege escalation (LPE) exploits against the Linux kernel in a mid-October run, each paired with a candidate patch, for $1,213,430 in API credits, or about $1,279 per exploit. OpenAI's own conclusion is blunt: for now, a low-privileged account on a system that is behind on patches "should be considered functionally equivalent to root access for a well-resourced threat actor."
What the pipeline did#
The pipeline chose security-relevant parts of the kernel, about 7% of its code, scanned them, reproduced each finding in an isolated virtual machine, tried to build an exploit, wrote a patch, and then handed everything to a human security researcher, OpenAI said. From 2,308 candidate findings it reproduced 1,724 vulnerabilities and produced proofs of concept for 291 information leaks. The exploits ran against a stock desktop distribution with standard protections switched on, not a weakened target.
OpenAI says the agents used 10 exploitation techniques that were new to it and to outside kernel experts it consulted, and it withheld the details "for safety reasons." The same harness running GPT-5.6 Sol in late July found 166 LPEs for $513,690; the newer model found more than five times as many novel LPEs despite three months of kernel hardening in between.
OpenAI also tested Google's hardened kernelCTF setup, where the pipeline produced 21 LPE exploits for about $32,936. For comparison, it says 26 confirmed zero-day LPEs had been accepted by kernelCTF in all of 2026.
Where the fixes stand#
Following the kernel community's guidance, OpenAI says it has sent hundreds of reports and patches for critical bugs and is working through the rest with partners in Patch the Planet, the open-source program it launched alongside GPT-5.5 Cyber in June. When the 10-day scan ended, 61 of the LPEs had already been patched upstream, 43 had staging patches and 299 were being discussed publicly, OpenAI said, which is why it urges everyone to keep systems updated.
The same day, OpenAI said GPT-5.6 Cyber is now available through Daybreak Red in AWS GovCloud (US-East) via Amazon Bedrock, for customers approved by both OpenAI and AWS. That extends the commercial Bedrock launch in August to U.S. government workloads. The post lands days after Anthropic described five exploit chains from an unreleased model.
Limits#
- Every number here is OpenAI's own. No maintainer, distribution or independent researcher has verified the 948 count, and coverage we found repeats OpenAI's post.
- OpenAI does not name the model, the Linux distribution or the kernel version, and has not said when the model will ship or to whom.
- The $18.9 million to $66.3 million "valuation" OpenAI cites applies past contest bounties to its exploits; it is not money anyone paid.
- OpenAI says an early run counted 20 of 200 "successful" exploits that actually used a different zero-day than the one assigned; it says it fixed that check, but the final figures are not externally audited.