OpenAI agents found 948 Linux root exploits in a single 30-hour run

OpenAI has revealed that its upcoming cyber-focused artificial intelligence model identified and patched 948 local privilege escalation vulnerabilities in the Linux kernel during a single 30-hour test run in mid-October 2026. The experiment, which cost over $1.2 million in API credits, demonstrates how autonomous AI agents can now accelerate both the discovery of critical security flaws and the generation of fixes at a scale previously impossible for human teams.

What happened

The research team deployed an agentic pipeline designed to audit the Linux kernel for security weaknesses reachable by low-privileged users. During the 30-hour session, the system analyzed 2,308 candidate findings and dynamically reproduced 1,724 vulnerabilities in a sandboxed environment. From these, it successfully generated working proof-of-concept exploits for 948 local privilege escalation (LPE) bugs and 291 information leak vulnerabilities. An LPE exploit allows a user with limited permissions to gain root access, effectively taking full control of the system.

This effort builds on earlier work from June 2026, where OpenAI’s GPT-5.5-Cyber model discovered 24 LPE exploits. The new pipeline automates the entire lifecycle of vulnerability management, moving from initial discovery to patch generation. The total cost for the run was $1,213,430, averaging $1,279 per confirmed LPE exploit. The agents also uncovered at least 10 novel exploitation techniques, including new methods for memory spraying and page grooming, which improve the reliability of attacks against the kernel.

The implications extend beyond this specific test. Data from the Google Threat Intelligence Group shows that monthly vulnerability disclosures more than doubled between January and August 2026, rising from 5,045 to 10,740. High-risk disclosures surged by 167% in the same period. The U.K. National Cyber Security Centre notes that while defensive AI tools can speed up patching, the widespread availability of these same tools lowers the barrier for attackers, making zero-day discovery faster and cheaper for adversaries.

How it works

The pipeline operates through five distinct stages managed by autonomous agents. First, in vulnerability discovery, agents scan the codebase to identify high-value targets that are security-relevant and accessible to unprivileged users. Next, during dynamic validation, the agents attempt to reproduce each potential vulnerability in a isolated sandbox, preserving runtime evidence to confirm the bug exists. This step filters out false positives before significant resources are spent on exploitation.

Once validated, the agents move to exploitability assessment, where they test if the vulnerability can be abused for information disclosure or privilege escalation. They then devise a proof-of-concept exploit. In the fourth stage, patch generation, the agents create code fixes for the identified issues. Finally, human security researchers review the exploits, patches, and severity ratings before reporting them to Linux maintainers. This human-in-the-loop step ensures quality and prevents erroneous patches from being deployed.

Key details

• The pipeline discovered 948 working local privilege escalation exploits and 291 information leak proofs-of-concept.
• The 30-hour run cost $1,213,430 in API credits, resulting in a cost of $1,279 per LPE exploit.
• Agents identified at least 10 novel Linux exploitation techniques, including new persistent elastic objects and data-only escalation methods.
• The system analyzed 2,308 candidate findings and dynamically reproduced 1,724 vulnerabilities in a sandbox.
• Monthly vulnerability disclosures rose from 5,045 in January 2026 to 10,740 in August 2026, according to Google Threat Intelligence Group.
• OpenAI states that unpatched systems with low-privileged access should now be considered functionally equivalent to having root access for well-resourced threat actors.

Why it matters

For software engineers and infrastructure leads, this shift means the window between vulnerability discovery and active exploitation is shrinking dramatically. Traditional patch cycles, which might take weeks or months, are no longer sufficient when AI agents can find and weaponize bugs in hours. The cost of entry for attackers is dropping, meaning even novice threats can leverage AI to find zero-days. As the NCSC warns, attackers are becoming less cautious about losing access because replacing it is cheap and easy.

However, this technology also offers a path for defenders. By integrating similar agentic pipelines into development workflows, teams can generate patches as quickly as vulnerabilities are found. OpenAI suggests that with human expert guidance, AI-generated patches can be robust and aligned with existing coding styles. The goal is to make it cost-prohibitive for bad actors to exploit software by closing holes faster than they can be found. For now, the priority must be rapid patching and automated testing to keep pace with AI-driven discovery.

What you can do

• Prioritize immediate patching for any Linux systems that allow low-privileged user access, treating them as high-risk.
• Integrate automated vulnerability scanning and patch generation tools into your CI/CD pipeline to reduce response time.
• Review and update internal security policies to account for faster vulnerability disclosure rates and shorter exploitation windows.
• Implement strict access controls to limit the number of users with low-privileged accounts on critical infrastructure.
• Monitor threat intelligence feeds for new AI-assisted exploitation techniques and novel attack vectors like those described by OpenAI.
• Conduct regular red-team exercises using AI-powered tools to identify weaknesses in your own systems before adversaries do.