OpenAI says a single run of an upcoming cybersecurity model produced 948 working Linux root exploits, each with a candidate patch, for $1,213,430 in API costs. In a report titled "A thousand fewer paths to root," published on 4 November 2026, the company put a price on automated exploit development, about $1,279 per distinct flaw. The numbers come from OpenAI itself, but they point to a real change in how fast software flaws can be found.
What OpenAI did
The mid-October run targeted the latest Linux kernel code. According to OpenAI's report and coverage of it, the pipeline worked in stages:
- Scouting: Scout agents picked 5,034 security-relevant targets covering about 2.8 million lines, roughly 7% of the kernel, for about $3,400.
- Analysis: Analyst agents produced 2,308 candidate findings in 12 hours, costing about $316,500.
- Validation: Agents reproduced 1,724 of them as real vulnerabilities, a 74.7% hit rate, for about $73,000.
- Exploits and patches: Exploit agents built 1,169 working privilege-escalation exploits, which mapped to 948 distinct vulnerabilities, each with its own patch, for about $776,000.
OpenAI says the agents used at least 10 exploitation techniques that outside experts had not seen before, and that human security researchers reviewed the results before they were reported to Linux maintainers.
The cost comparison
OpenAI compared the $1,279 average per flaw with past prizes at the Pwn2Own and TyphoonPWN hacking contests, which have paid $20,000 and $70,000 for a Linux kernel privilege-escalation exploit. On those rates the 948 exploits would be worth $18.9 million to $66.3 million, or 16 to 55 times the run's cost. That is a hypothetical comparison made by OpenAI, and the company describes its cost figures as an upper bound that better engineering could reduce.
What this does and does not mean
These are local privilege-escalation flaws. They let someone who already has a low-privilege account on a machine gain full administrator, or root, control. They are not remote attacks that work over the internet by themselves. Even so, OpenAI warns that a low-privilege account on an unpatched system should now be treated as functionally equivalent to root access for a well-funded attacker, because the same capability could compress the time between a code change and a working exploit.
Disclosure and limits
OpenAI says it is not publishing the technical details for safety reasons. Following advice from the Linux kernel community, it started with the most critical vulnerabilities and is sending the rest gradually, working with partners from its Patch the Planet initiative. The model used is unreleased, so the results cannot be independently reproduced yet, and the count of 948 is OpenAI's own tally of unique flaws, based on distinct patches.
What to do
- Apply Linux kernel and distribution security updates promptly, especially on shared servers and systems where many people have accounts.
- Limit who has shell access, and remove accounts that are no longer needed.
- Treat container and sandbox boundaries as something to harden, since a kernel flaw can undermine them.
For defenders, the result cuts both ways. The same pipeline that finds flaws for attackers can find and patch them first, and the bottleneck may shift from finding bugs to fixing them quickly.
Sources: OpenAI, "A thousand fewer paths to root" (4 November 2026), as reported by AI Weekly, RuntimeWire, NewsBytes, ShiftDelete, BigGo Finance and ai-tldr.
