OpenAI pipeline generates 948 Linux root exploits for $1.2M in API credits
- OpenAI's pipeline scanned the Linux kernel for 2,308 candidate findings, dynamically reproduced 1,724 of them, and built 948 working local privilege escalation exploits plus 291 information leak PoCs, at a total cost of $1,213,430 in API credits, or $1,279 per LPE.
- Priced against the historical Pwn2Own ($20,000) and TyphoonPWN ($70,000) Linux kernel LPE bounties, the 948 exploits are worth a hypothetical $18.9M to $66.3M, or 16x to 55x the pipeline's cost.
- On an adapted kernelCTF environment with reduced attack surface, the pipeline produced 25 information leak PoCs and 21 LPE exploits for about $32,936, while kernelCTF accepted only 26 confirmed 0-day LPEs in all of 2026 to date.
- The run found at least 10 novel Linux exploitation techniques and over 5x more novel LPEs than a GPT-5.6 Sol run in July, even with three months of kernel hardening in between.
- OpenAI reports findings and candidate patches to Linux maintainers through Patch the Planet, starting with critical vulnerabilities; within 10 days of the scan, over 100 Linux LPEs had been patched or had staging patches, with roughly 300 more under public discussion.
Hacker News opinions
This is my team's work. Happy to answer questions or take them back to the team, and we're hiring.
It's bittersweet, like finding asbestos in your house. I'd rather know than not, but I'd much rather it wasn't there. Any plan for continuous scanning of new PRs as they land?
I wish these posts broke down the categories of defects behind the bugs and what would prevent them. A system that admits thousands of these is flawed in a few fundamental ways.
948 results is a lot to review in a timely manner. How do you decide when to trust the model and when a human has to look at it?
The number is impressive to the point I'm skeptical. Are these distinct vulnerabilities or a lot of closely related ones? And are they kernel bugs or bugs in things running as root?
I'd want a breakdown by age of the code and by subsystem: drivers, networking, core.
The Linux Foundation has Sashiko running Google models over every kernel commit. How many of your 948 did Sashiko already catch?
Do you have a list of the publicly discussed ones? I used to handle nonpublic kernel security reports and would love a small sample. And don't stop at LPEs and infoleaks, semantic bugs are interesting too.
Imagine if open source projects beyond the Linux Foundation could throw $1.2M of API credits at security.
That applies to anything capital-gated. Telescope time, atom smashers, quantum labs. Frontier models on huge datacenters are a new kind of captured commons.
OpenAI says a low-priv account on an unpatched machine equals root for a well-resourced attacker. Buries head in sand. That's been true for decades, only the frequency changed.
LPEs are frequent enough that nobody sandboxes malicious software with low-priv accounts anymore. They just separate well-behaved users.
How did we get to a thousand privilege escalation exploits in one kernel? Is it 40+ million lines of memory-unsafe code, or a broken security model?
Multi-user died in the 90s, privilege escalation has been table stakes since then. The front door was always unlocked; what changed is that exploit discovery economics now favor the attackers.
Posted an invite to ask questions, then answered nothing. Classic.