OpenAI says cyber model found 948 Linux kernel privilege escalation flaws

OpenAI says an upcoming cyber model helped identify 948 distinct privilege escalation flaws in the Linux kernel and that its team built patches for them. The company’s report describes a five-stage pipeline covering vulnerability discovery, dynamic validation, exploitability assessment, patch generation and human review.

In the first 12 hours, AI agents produced more than 2,300 potential findings. Further agents generated proof-of-concept code and tested candidates. The 30-hour process reproduced 1,724 findings, about 75% of the initial candidates. OpenAI said roughly 3% of the remaining findings came from incorrect model analysis.

The agents then ranked vulnerabilities by impact, producing 1,169 working local privilege escalation exploits and 291 proof-of-concept demonstrations of information disclosure. After removing duplicates and false positives, researchers counted 948 distinct escalation flaws. OpenAI said researchers reviewed the generated exploits and patches before contacting Linux kernel maintainers.

By the end of the analysis, maintainers had already patched 61 of the identified flaws, while 43 had staging patches. Just under 300 were being publicly reported and discussed. The team also examined 166 exploitation techniques and identified 10 it said were previously undocumented. OpenAI withheld details until the flaws are fixed.

The report says the work cost about $1.21 million in API credits. It does not name the upcoming model, and compares its results with GPT 5.5-Cyber, which the report says developed 24 local privilege escalation exploits.