OpenAI unveils artificial intelligence for cybersecurity, confined to an isolated environment

OpenAI has unveiled GPT-6 Cyber, an artificial intelligence model specialized in cybersecurity. The news currently comes from a single source (OpenAI statement, reported by Red Hot Cyber); no independent confirmation is currently available in the material gathered.

The model is designed to identify vulnerabilities in software, develop and run tests on exploits — that is, programs that exploit a security flaw to demonstrate its danger — and prepare the relevant corrective patches. According to reports on tests conducted by the company, GPT-6 Cyber identified 81 unknown vulnerabilities in the Firefox browser, reproduced 1,724 known flaws in the Linux kernel, and developed 948 working exploits.

According to the company’s announcement, these results represent a clear leap forward compared to the previous version, GPT-5.6 Cyber, released in August: the data, however, comes from internal tests conducted by OpenAI itself, and no verifications carried out by parties outside the company appear in the material gathered.

Precisely because of the dual nature of a tool capable of both defending and attacking computer systems, OpenAI has chosen not to release the model in open form: GPT-6 Cyber remains confined to an isolated security environment, a sort of digital enclosure that prevents the model from operating outside the controlled context in which it was designed. It is a choice that the company itself describes as necessary to prevent the same capabilities used to discover flaws from being exploited to create new ones without supervision.

The case fits into a broader debate on the dual use of artificial intelligence in the field of computer security, where the same technology that identifies a vulnerability to protect a system can, in theory, be used to attack it. The material gathered contains no indication of a public availability date for the model outside the controlled environment described by the company.