OpenAI sets AI hunting for flaws in critical systems: what changes with GPT-6 Cyber

OpenAI has made available Gpt-6 Cyber, a model developed specifically to identify critical vulnerabilities in internet-connected corporate systems, according to ANSA. The model has already been used to analyze the source code of Linux, the operating system underlying most of the world’s servers and digital infrastructure, tracking down hundreds of potential issues.

The news is also confirmed by the Chinese portal Sina, which reports the same essential elements of the release, offering a second source independent of the first.

The launch does not come in isolation. In September OpenAI had announced a one-billion-dollar investment earmarked for critical infrastructure security: the new model fits into that strategy, shifting the application of artificial intelligence from generic text toward the systematic search for errors in the code that manages power grids, industrial plants and connected corporate systems.

Compared to the previous version, according to reports, the model has improved its ability to stay within the authorized perimeter during scans, avoiding interaction with external services not covered by the test. This is a far from minor technical detail: an automated system searching for security flaws on real networks must be able to operate without producing side effects on third-party infrastructure, a risk inherent to any large-scale scanning tool.

The question of control over the use of a tool of this kind remains open, and is not resolved by the available sources: a technology capable of uncovering vulnerabilities in critical systems is by definition dual-use, useful both to those who must defend and to those who might seek to exploit them. Neither of the two sources consulted provides details on who will have access to the model or on any restrictions on its use, a point that remains to be verified in future developments.