OpenAI scans the Linux kernel, discovers 948 vulnerabilities, and urges IT departments towards load isolation

OpenAI has entrusted the Linux kernel to a team of vulnerability research agents, resulting in 948 privilege escalation exploits and an equal number of candidate patches, averaging $1,279 in API credits per exploit. The company concludes that an ordinary user account on a system outdated with patches now equates to administrator access.

Researching vulnerabilities within the extensive Linux kernel—comprising over 40 million lines of code that power everything from Android devices to cloud infrastructure—has traditionally been a painstaking process, carried out by researchers on a case-by-case basis. However, OpenAI, the developer behind ChatGPT, announced on November 4 that it conducted a comprehensive industrial scan of the kernel in mid-October using a cybersecurity model that has not yet been made public.

The standout figure from OpenAI’s scan is the identification of 948 distinct vulnerabilities that allow low-privilege users to gain root access on common desktop distributions, marking a significant scale-up in vulnerability discovery. Back in June, during the Patch the Planet initiative, the GPT-5.5 Cyber model discovered 24 similar exploits. The scan conducted in October utilized $1,213,430 worth of API credits, a figure that OpenAI refers to as a ceiling that optimization could potentially reduce.

A Streamlined Chain of Agents for Research

The vulnerability discovery process is organized into five steps, from initial detection to human review. First, exploratory agents focus on sections of the kernel that are accessible to low-privilege users, identifying 5,034 targets within approximately 2.8 million lines of code—around 7% of the total—at a cost of $3,426. Subsequently, analysis agents sift through this data to generate 2,308 candidate findings within just twelve hours, costing $316,549.

Each finding must then be validated, which involves replaying each candidate within a virtual machine equipped with a memory error detector. This rigorous testing confirmed 1,724 vulnerabilities over a span of thirty hours, resulting in a 74.7% replication success rate. OpenAI indicated that only 3.1% of the candidates were based on false analyses, with the remainder requiring elevated privileges or activation conditions that were too rare for the allocated time.

Next comes the most resource-intensive phase, enlisting exploit development agents at a cost of $775,929 in credits. These agents stress-test the candidates on machines configured with standard distribution protections, including random kernel address space layout. This phase yielded 1,169 functional exploits, distilled down to 948 distinct vulnerabilities after examining variations. On kernelCTF, a reduced-surface version of the kernel used for competitions, the chain produced 21 exploits for $32,936, along with the acceptance of 26 previously unknown vulnerabilities since the beginning of 2026. OpenAI noted that its agents employed at least ten new exploitation techniques.

Candidate Patches Face Maintenance Review Bottlenecks

The figure of 948 relates to proposed patches, not closed vulnerabilities. OpenAI sends these to kernel maintainers, prioritizing critical vulnerabilities based on advice from the Linux community and partners in the Patch the Planet initiative. However, OpenAI has not disclosed how many of these patches have been integrated into the official kernel.

This shift indicates that discovery isn’t the bottleneck—it is the review process that takes precedence. During the ten days of the scanning operation, over 100 privilege escalations were either fixed or placed on hold by other contributors, with approximately 300 additional discussions occurring publicly. OpenAI argues that the automatic generation of patches is becoming essential to keep pace with discoveries and to shorten the exploitation window.

To contextualize the value of its findings, OpenAI compared its historical figures from the Pwn2Own and TyphoonPWN competitions, which award between $20,000 and $70,000 per exploit. Based on this, the 948 exploits could be worth between $18.9 million and $66.3 million. The exact model used for this research remains unnamed, and the tested distribution is only described as commonly used, which complicates comparisons with other laboratory efforts.

Local Accounts Now Equivalent to Root Access

The generational leap in vulnerability discovery highlights a crucial warning. The mid-October model uncovered over five times the number of new vulnerabilities compared to a scan from GPT-5.6 Sol in July, despite three months’ worth of kernel patches between the two assessments. OpenAI deduces a key rule for defenders: in the short to medium term, a low-privilege account on a system lacking the latest patches should be treated as equivalent to root access by an attacker with sufficient resources.

Privilege escalations presuppose initial system access through a compromised account, vulnerable application, or shared container. This perspective shifts focus towards two leverage points within the IT department’s control: the timely application of kernel patches delivered by distributions and the compartmentalization of workloads. OpenAI cites Firecracker micro-VMs and gVisor sandboxing as methods to limit the potential impact of an exploited vulnerability.

Isolation and Patch Cadence: New Priorities for IT Departments

For IT departments, OpenAI’s scanning results may change priority dynamics more than they alter tool sets. While enterprise distributions will continue to deliver patches, the volume of these patches is expected to grow with each model generation, and shared servers—whether housing multiple accounts or containers—will increasingly be judged by their isolation capabilities in addition to their software versions. Kernel updates, which are often delayed due to the need for reboots, are moving up the priority list.

OpenAI anticipates that within a few years, advanced models will identify every existing vulnerability in software, with less capable models understanding how to exploit them once disclosed. The same capabilities offered to defenders through OpenAI’s Daybreak platform will, in turn, compress the timeline between patch publication and its potential exploitation.

Scroll to Top