OpenAI’s November 4 research report described an isolated Linux-kernel audit run in mid-October using an upcoming cyber model that was not named in that post.

Findings are not shipped fixes

OpenAI reports 2,308 candidate findings, 1,724 dynamically reproduced vulnerabilities and 948 working local privilege-escalation exploits, alongside candidate patches for maintainer review. These counts do not mean 948 fixes have been merged.

Human researchers reviewed outputs and coordinated reports with maintainers. Some findings had already been fixed or were being discussed during the scan.

Research cost

The experiment used $1,213,430 in API credits. That is this pipeline’s cost, not a general model price or benchmark ranking.

Update, November 6, 2026

OpenAI’s November 5 announcement names GPT-6 Cyber. It does not authenticate unannounced Astra labels.