New artificial intelligence models are increasingly unpredictable, as shown by recent cases of AI systems slipping out of human control and breaching entire systems in a matter of minutes. Those incidents have underscored how important it is becoming to have technologies capable of countering AI-driven attacks.

Microsoft was among the first companies to move in this direction. On July 28 it expanded its lineup of security tools with new products designed to help companies protect their computer networks. The idea is to use the same tools that are exploited to attack, turning the power inadvertently placed in the hands of hackers back against them.

Microsoft’s new AI model

This is where MAI-Cyber-1-Flash comes in, the new Microsoft artificial intelligence model trained for cybersecurity. The system, the company says, acquired its skills by analyzing decades of data gathered in response to the hacking attacks suffered by its customers.

The database used for training is enormous, considering that Microsoft owns the Windows operating system, Outlook email, and Azure cloud computing, three of the most widely used services in the world and among the most frequent targets of cyberattacks.

Unlike other tech giants leading the AI sector, Microsoft chose not to share the new model with independent beta testers before release. On the contrary, confident in the results it had achieved, the company said it expects MAI-Cyber-1-Flash to place first right away in a standard benchmark test called CyberGYM, surpassing the solutions offered by OpenAI and Anthropic.

Inside MAI-Cyber-1-Flash

The cost of running MAI-Cyber-1-Flash is roughly half that of other technologies on the market, according to Microsoft. The system can reportedly handle 90% of requests effectively, meaning the more expensive models would come into play only in the remaining 10% of cases.

But the real breakthrough is something else. Alongside MAI-Cyber-1-Flash comes Project Perception, a tool capable of turning different AI models into teams of agents designed to detect and repair network vulnerabilities. In practice, these assistants will be able to use other software to carry out tasks autonomously. Some of them should even be able to imitate hackers.

In this way, it will be easier to intercept potential system flaws well in advance, allowing fixes to be made before a bad actor can exploit them.

Editor’s note

This article was originally published in Italian on money.it by Pasquale Conte on July 29, 2026 as «Microsoft lancia il suo primo modello AI per la cybersecurity. Sfida ad Anthropic, Google e OpenAI». It has been translated and adapted for an international audience by the Money.it International desk.