OpenAI says an internal AI pipeline found, reproduced and turned Linux kernel bugs into 948 distinct working local privilege escalation exploits in a single October campaign, at a cost of about $1.2 million in API spend. The company published the results on 4 November in a report titled "A thousand fewer paths to root", and says every finding is being routed to kernel maintainers with a candidate patch.
What the pipeline did
According to OpenAI's report as summarised by CSO Online and Startup Fortune, the system started with 2,308 candidate findings in the Linux kernel. Of these, 1,724 were reproduced as real vulnerabilities, and 948 went on to become validated exploits that give an attacker with an existing foothold on a machine full root access. The pipeline automates discovery, dynamic validation, exploit development and patch generation, with human researchers reviewing each root-cause analysis, proof-of-concept and proposed fix before anything is sent to a maintainer.
The total bill came to $1,213,430, which works out to roughly $1,279 per exploit. OpenAI also estimates the exploits' market value at between $18.9 million and $66.3 million, based on the $20,000 to $70,000 bounties that contests such as Pwn2Own and TyphoonPWN have paid for this class of bug. That comparison is OpenAI's own framing, not a record of actual payouts, and it does not mean every finding would qualify for a reward.
The company says the run produced about five times as many unique privilege escalation bugs as an earlier experiment with a model it calls GPT-5.6 Sol, even though the kernel had gained three more months of hardening in between. An analysis of 166 of the exploits turned up at least 10 exploitation techniques that had not been seen before.
Where the patches stand
The headline number is not 948 fixes. When the campaign was written up, 61 of the vulnerabilities had already been patched upstream, 43 more had patches staged, and 299 were under public discussion with maintainers. OpenAI says it has submitted hundreds of reports and patches, but it has not published a full breakdown of what has happened to every finding.
The work sits inside Patch the Planet, an initiative OpenAI launched in June under its Daybreak programme with partners including Trail of Bits and HackerOne. At that launch, an earlier model produced 24 privilege escalation proofs of concept. Startup Fortune reports that engineers at Trail of Bits sign off on findings before they go to maintainers, though OpenAI's own page was not available to us to confirm that detail.
OpenAI is open about the weak points. Early AI-written patches sometimes added unnecessary code, introduced performance trade-offs or missed related variants of the same bug. In one early experiment, 20 of 200 exploits that looked successful had gained root through a different zero-day than the one the agent was assigned, so the team added checks to confirm the agent had done the job it was given.
Why the caveats matter
All of this was run by OpenAI in controlled, company-run experiments, and the cost and result figures come from its own account. We have not seen independent reproduction of them. The attacks also need local access, so they are a step in a longer chain rather than a way to break into a machine from outside.
Security analysts quoted by CSO Online point to the gap between discovery and deployment. IDC's Sakshi Grover said the risk is "less the 948 than the length of the window" before fixes reach production. An upstream patch does not mean a fix has reached the enterprise distribution kernels that most organisations actually run, and distributors have to backport changes on their own schedules. Analysts advise starting with multi-tenant hosts, CI and CD runners, shared bastion servers, container platforms and developer fleets, and checking vendor backport timelines.
There is also a human bottleneck. Validating AI-generated exploits and writing reliable root-cause explanations takes people time, which slows reporting. Security Brigade's Yash Kadakia warned that achieving a result and explaining it correctly are two different things, and urged buyers of such tools to demand reproducible exploits and independent validation.
The wider picture
OpenAI describes the work as defensive, aimed at fixing bugs before attackers find them. The same cost curve cuts both ways, though: if one team can generate nearly a thousand root exploits for about $1.2 million, the economics of kernel bug hunting have shifted for everyone.
For Indian organisations that run Linux across data centres, cloud estates and developer machines, the practical message is about patch speed. Teams that depend on distribution kernels from Red Hat, Ubuntu, SUSE or a cloud provider should watch for advisories and apply them quickly rather than waiting for a routine cycle.
What to watch next
OpenAI says it plans to extend the approach to other critical open-source projects through Patch the Planet. Watch for how many of the 299 bugs under discussion reach upstream releases, how quickly major distributions backport them, and whether outside researchers can verify the exploit counts. A published remediation breakdown from OpenAI would also show how much of the 948 turns into shipped fixes.
Published 6 November 2026. Gadgets365 will update this article as more information becomes available.
