OpenAI’s new ‘GPT‑6 Cyber’ model exposes nearly 1,000 Linux flaws, stoking security fears

On November 5, OpenAI quietly began rolling out a new security‑focused model called GPT‑6 Cyber to a small group of verified customers. According to reporting based on OpenAI’s internal tests, the system was able to probe the Linux operating system—core infrastructure for servers, cloud platforms and many phones—and uncover close to 1,000 previously unknown software vulnerabilities.

The breakthrough highlights the double‑edged nature of frontier AI. For defenders, a model that can automatically generate exploits and test patches at scale could dramatically speed up the process of discovering and fixing bugs before criminals get to them. But in the wrong hands, the same capability could be turned into an efficient weapon for mass cyber‑attacks across critical infrastructure.

OpenAI itself has assigned GPT‑6 Cyber its highest internal risk rating, “critical,” and is limiting access while it works through safety and governance questions. The launch adds urgency to ongoing debates in governments and standards bodies over how to regulate powerful “offensive‑capable” AI models—systems that can both secure and compromise the digital systems modern economies rely on.