OpenAI's Cyber Model Found 948 Linux Kernel Flaws, Then Patched Them

What OpenAI's cyber model actually found in the Linux kernel

OpenAI ran a five-stage automated pipeline against the Linux kernel and ended up with 948 distinct privilege-escalation vulnerabilities, along with patches for them.

The Linux kernel is more than 40 million lines of code and sits under Android, cloud infrastructure, supercomputers and plenty of other systems. OpenAI picked it precisely because of that scale and importance, and the company says the goal was to see how well frontier models could find real vulnerabilities in a codebase that size.

The pipeline moved through discovery, dynamic validation, exploitability assessment, patch generation and human review. In the first stage, AI agents scanned security-relevant parts of the kernel and produced more than 2,300 candidates in 12 hours. Additional agents then wrote proof-of-concept code to test each one.

Validation took about 30 hours and successfully reproduced 1,724 findings, roughly 75% of the initial candidates. Only 3% of the remaining findings came from incorrect model analysis, according to OpenAI.

From there, the agents ranked flaws by impact and developed 1,169 working local privilege-escalation exploits plus 291 proofs of concept for information disclosure. After removing duplicates and false positives, the team arrived at 948 distinct escalation flaws and built patches for them.

The $1.21 million price tag and the launch hint

The whole exercise cost roughly $1.21 million in API credits, and OpenAI never names the model behind it, calling it only an 'upcoming cyber model.'

That unnamed model is compared in the report to GPT 5.5-Cyber, which 'developed 24 local privilege escalation exploits.' The contrast is the clearest signal yet that OpenAI is preparing a security-focused release, and the report reads partly like a preview of what that model can do.

The cost matters too. $1.21 million in API credits is not a hobbyist budget, but it is small next to what a serious bug-bounty program or a dedicated security team might spend chasing the same class of flaws across a kernel this large. If the numbers hold up, AI-assisted auditing starts to look economically sensible at scale.

How many of those flaws were actually new?

Not all 948 were brand-new discoveries, and OpenAI is upfront about that.

Linux maintainers had already patched 61 of the identified privilege-escalation flaws by the time the analysis finished, and another 43 had staging patches in progress. Just under 300, roughly a third, were already being reported and discussed publicly.

OpenAI says its researchers reviewed every generated exploit and patch before contacting maintainers, then ranked the flaws by priority following guidance from the Linux community. Some have already been patched; the rest are still being addressed.

The team also examined 166 exploitation techniques and found 10 that were previously undocumented, including kernel memory manipulation, page grooming and privilege escalation through data-only attacks. OpenAI is holding back details on those until they are fixed.

What this means for AI security claims

A model that can find, validate and patch real kernel bugs changes the economics of security research, but it also raises questions about disclosure and trust.

OpenAI's report is a strong demonstration, and the fact that maintainers are already working through the findings gives it credibility that a pure benchmark score would not. The company also did the right thing by coordinating rather than dumping exploits publicly.

Still, the report is marketing as much as research. The model has no name, no release date and no independent audit in the document. The real test will be whether the patches hold up and whether the unnamed cyber model ships with the same capabilities for customers. For now, the Linux kernel is a little safer, and OpenAI has a very detailed teaser for whatever it launches next.

Availability and pricing

OpenAI has not announced a name, release date or price for the cyber model described in the report. The Linux kernel findings are already being handled through the usual maintainer process, so users of Linux-based systems do not need to take any action beyond normal updates.

When the model does launch, expect it to be offered through OpenAI's API and possibly higher-tier plans, but nothing has been confirmed.

OpenAI has not said when the cyber model will launch or what it will cost, but the Linux kernel report makes the direction clear. The company is building AI that does not just flag bugs but proves them and writes fixes. Whether that makes software safer or just shifts the workload to maintainers is the open question.