OpenAI says an automated system built from AI agents has uncovered hundreds of serious security weaknesses in the Linux kernel, in a research run that shows how quickly artificial intelligence is changing both the offence and defence sides of cybersecurity.
In a research post published on Wednesday, the company described a prototype pipeline that used an upcoming cybersecurity-focused model. In a run in mid-October that took about 30 hours, the system analysed 2,308 candidate findings in the Linux kernel, reproduced 1,724 as real vulnerabilities, and ended with 948 validated flaws that would allow a local user to escalate their access to full “root” control of a machine.
What the system did
The pipeline works in stages, with different AI agents handling different jobs. Scout agents choose which parts of the kernel are worth examining. Analyst agents read the code looking for flaws. Triage agents then try to reproduce each suspected bug in an isolated virtual machine. Further agents test whether a confirmed flaw could be abused, and a final stage writes candidate patches.
According to OpenAI, the run produced working proofs of concept for 948 distinct “local privilege escalation” problems, plus 291 cases where information could leak. The company says the agents also used at least 10 previously undocumented exploitation techniques. The tests were carried out on a standard Linux distribution inside controlled environments.
The cost was about $1.21 million in computing credits, or roughly $1,279 for each validated flaw. For comparison, an earlier version of the system, built on the GPT-5.6 Sol model, found 166 such flaws in July at a cost of around $514,000. The October run found more than five times as many, despite three months of additional hardening of the kernel itself.
Why this is significant
Local privilege escalation does not mean someone can break into a computer from the internet. An attacker generally needs a foothold first, such as a low-privilege account or a compromised application. But these flaws are valuable precisely because they turn a small foothold into complete control, and they are a common step in real-world attacks on servers, cloud systems and devices.
The scale is what stands out. Security researchers have long found Linux bugs one at a time, often over weeks or months. The idea that software can find, confirm and propose fixes for hundreds in a day and a half suggests that the economics of vulnerability research are shifting. The same capability that helps defenders could, in the wrong hands, help attackers, which is why OpenAI is describing the work as defensive research and stressing its disclosure process.
What happens to the bugs
It is important to be clear about what the number means. The 948 figure refers to validated flaws with candidate patches produced by the pipeline, not 948 fixes that have already been shipped to users. OpenAI says it has begun sending reports and patches to Linux kernel maintainers and is working through an initiative called Patch the Planet to get further fixes to the right people.
Kernel developers will still need to review each proposed change. Maintainers of open-source projects are often volunteers, and a sudden flood of reports, even good ones, can strain their time. That tension, between faster discovery and limited capacity to fix, is likely to be a theme of the next few years.
What it means for businesses and users
For most people, nothing needs to change today. Operating-system updates are the main protection, and users and administrators should keep Linux systems, phones, routers and cloud servers patched as vendors release fixes. Companies that run Linux at scale will want to watch for security advisories linked to this work and follow their usual patching schedules.
For security teams, the research is a warning about pace. If AI systems can find large numbers of flaws cheaply, the time between a bug existing and it being known could shrink. Many experts argue the answer is to use the same tools defensively, scanning software before attackers do and automating the writing and testing of patches.
What to watch next
Key questions include how the Linux community responds, whether other AI labs publish comparable results, and how regulators and companies handle the risk that powerful cyber-capable models could be misused. OpenAI has described the current model as upcoming rather than publicly released.
The finding adds to a wider debate about how openly such capabilities should be shared. For now, the message from the research is that artificial intelligence has become a serious tool for finding software flaws, and that defenders will need to move faster to keep up.
