OpenAI’s automated security workflow dynamically reproduced 1,724 Linux kernel vulnerabilities and generated 948 privilege-escalation exploits, consuming approximately $1.21 million in API credits.

The results came from an October 2026 test described in a November 4 security report. According to the supplied summaries, OpenAI used an upcoming cybersecurity model to examine 2,308 candidate Linux-kernel vulnerabilities and dynamically reproduce 1,724. The workflow comprised discovery, dynamic validation, exploitability assessment, candidate-patch generation, and human review.

The 948 exploits concerned local privilege escalation. These demonstrations allowed an ordinary local user to obtain root-level system privileges; the report does not characterize them as 948 remote compromises. OpenAI also reported 291 proof-of-concept programs for information-disclosure vulnerabilities and use of 10 new exploitation techniques that third-party kernel-vulnerability experts reportedly assessed as novel.

The reported resource cost was approximately $1.21 million in API credits. One summary identifies this as the full-process cost of the second scan, while another gives an average of approximately $1,279 per privilege-escalation exploit. These are API-credit figures, not disclosed GPU-hours, electricity consumption, or a complete accounting of research and human-review costs.

Patch submission and review were ongoing. OpenAI said it had prioritized submitting hundreds of reports and patches involving serious vulnerabilities in accordance with Linux-kernel community guidance, with candidate patches subject to human security review. A separate summary estimated hypothetical bounty-equivalent value of $18.9–66.3 million, or 16–55 times the workflow cost; that estimate is not realized bounty income, and the reports do not establish that every patch had been accepted or deployed.