OpenAI said an unreleased cybersecurity model uncovered 948 distinct local privilege-escalation vulnerabilities in the Linux kernel after running an automated pipeline that covered discovery, validation and exploit development. The company said the effort cost about $1.2 million in total.

In the account cited by MarsBit, the Linux kernel was described as core infrastructure behind Android, a large share of the world’s cloud servers and nearly all supercomputers. The report said the kernel contains more than 40 million lines of code and has been built over decades by top developers. OpenAI said that before AI tools entered this field, elite human security researchers would find about 20 similar bugs in the Linux kernel in a year, while this pipeline completed its work over a 10-day scanning and review window.
How the pipeline was structured
OpenAI broke the system into three stages: discovery and analysis, validation and reproduction, and exploitability assessment. Different AI agents handled each part.
Discovery and analysis
At the first stage, scout agents read through the full Linux kernel codebase and identified areas tied to security and likely attack surfaces. OpenAI said they narrowed more than 40 million lines of code to 2.8 million lines of high-value targets, roughly 7% of the total, at an API cost of $3,426.

Analyst agents then scanned those 2.8 million lines. OpenAI said they produced 2,308 suspected vulnerability reports in 12 hours, with spending of $316,000.
Validation and reproduction
The second stage was handled by a triage agent. For each suspected issue, it automatically wrote a proof of concept, or PoC, and ran it inside an isolated virtual machine to test whether it could trigger a crash.
OpenAI said the system reproduced 1,724 kernel-crash vulnerabilities within 30 hours. It reported a hit rate of 74.7%, while outright false positives caused by model hallucination came in at 3.1%.

Exploitability assessment
Once a bug was confirmed, ExploitDev agents attempted to bypass protections including SMEP, SMAP and KASLR to determine whether the flaw could be used to gain root privileges.
OpenAI said that under strict conditions, the ExploitDev agents produced 291 information-leak PoCs and 1,169 successful privilege-escalation exploits that obtained root access. After merging related variants, the final result was 948 independent local privilege-escalation flaws. Based on the figures in the report, the average cost to find and break a single flaw was about $1,279. For a complete working root exploit, the average API spend was about $359.
OpenAI said the model developed new attack methods
The report said the security team sampled 166 exploits written by the agents and found that the AI had independently discovered and used at least 10 new Linux kernel exploitation techniques.

- Two persistent elastic objects for precise heap spraying in kernel caches.
- Two new page-grooming techniques designed to improve privilege-escalation reliability.
- Five data-only attack methods using arbitrary write or direct data-structure corruption.
- One new arbitrary type-confusion technique.
OpenAI said it asked experienced external kernel experts to review those exploitation methods, and the response was that they had not seen them before.
Some successful exploits used different bugs than assigned
The article also described an early testing episode. At that stage, the security team instructed ExploitDev agents to write exploits for specific assigned vulnerabilities, with success defined as obtaining a root shell.
But in 200 successful cases, manual review found that 20 of the exploits did not use the assigned vulnerability at all. OpenAI said the agents kept working inside the constrained target environment, decided the assigned bug was too difficult to exploit, then found another previously unknown 0-day in the system and used that path to reach root.

The report presented that as a sign of tactical autonomy that exceeded what human researchers had expected from large models in offensive and defensive security work.
Comparison with a July run
OpenAI also compared the latest result with an earlier scan. It said that in July it ran GPT-5.6 Sol on the same task, spent more than $500,000 and found 166 vulnerabilities. Three months later, after switching to a stronger cybersecurity model, the count rose to 948 distinct local privilege-escalation flaws, more than five times the earlier result.
The original article added that, based on historical payouts at contests such as Pwn2Own, the 948 vulnerabilities could imply a value range of about $18.9 million to $66.3 million.

Patch the Planet and open-source remediation
The report cited researcher Eric Wallace, who said on social media that OpenAI is using the "Patch the Planet" project to direct large amounts of frontier compute to the open-source ecosystem on a pro bono basis for vulnerability discovery and remediation.
Wallace wrote: "We are now using a large amount of frontier compute, free of charge, to find and patch vulnerabilities in open-source codebases."
Beyond the Linux kernel, OpenAI said the same pipeline has been connected to Firefox and Chrome. The report said it recently helped Firefox fix more than 40 vulnerabilities.

Linux maintainers were patching during the review window
The article said that during the 10-day kernel scanning and manual review window, the Linux community was also moving on fixes. While the AI system was still validating more than 1,000 issues, Linux maintainers and developers had already patched 61 related local privilege-escalation issues upstream, had 43 patches in staging tests, and were publicly discussing nearly 300 more items on community mailing lists.
The original source cited in the MarsBit piece was OpenAI’s post, "A thousand fewer paths to root." MarsBit also said the article was republished from the WeChat account Quantum Bit, written by Jay.


