OpenAI said on Wednesday that an AI pipeline built 948 working Linux root exploits in about 30 hours. The run cost $1,213,430 in API credits, which works out to about $1,279 per exploit.
The company used an upcoming AI model focused on cybersecurity. The pipeline also wrote patches for the flaws it found, and security researchers reviewed the work before reports went to maintainers.
The test took place in mid-October 2026 on a commonly used Linux distribution. OpenAI did not name the distribution in its statement.
What the Pipeline Found
The system flagged 2,308 candidate findings in the Linux kernel. It then reproduced 1,724 of them as real vulnerabilities in a sandboxed setting.
From those, it built 948 local privilege escalation exploits. These let a user with a basic account gain root access, which means full control of the machine.
The pipeline also created proof-of-concept exploits for 291 information leak flaws. These are bugs that can expose data that should stay private.
OpenAI said the agents found and used at least 10 new Linux exploitation techniques. These include five data-only methods to gain higher access, two page grooming methods to make exploits more reliable, and two new ways to spray objects inside the kernel cache.
The tenth technique involves arbitrary type confusion. OpenAI said all 10 are new to the best of its knowledge.
The work builds on 24 exploits found by OpenAI’s GPT-5.5-Cyber model in June 2026. This time, the company turned the process into a five-stage system run by AI agents.
The stages are vulnerability discovery, dynamic validation, exploitability assessment, patch generation, and human review. In the last stage, researchers check the exploit, the patch, and the proposed severity rating.
Vulnerability Reports Are Rising Fast
The release comes as outside groups track a sharp rise in disclosed flaws. Google Threat Intelligence Group data shows monthly disclosures rose from 5,045 in January 2026 to 10,740 in August 2026.
High-risk disclosures grew from 131 to 350 over the same period. That is a 167% increase.
Last month, VulnCheck labeled AI-assisted vulnerability discovery a “growing risk category.” The firm said these tools let novice hackers find software flaws with little or no technical skill.
On Tuesday, the U.K. National Cyber Security Centre said defensive AI tools have shortened the time it takes to patch flaws. It also warned that many systems remain exposed if the tools are not set up correctly and widely.
The NCSC said attackers now show less caution about losing access to a system, since the cost to replace that access is low. It added that organizations will need to update their patching and testing processes.
OpenAI said automated patching must keep pace with AI-driven discovery. The company said AI-written patches can be correct and match existing coding styles when guided by human experts.
OpenAI also gave a direct warning to system owners. “For the near and medium term, having access to a low-privileged account on an operating system that isn’t up-to-date on its latest available patches should be considered functionally equivalent to root access for a well-resourced threat actor,” the company said.
