Short: attackers use AI, so defenders also put AI at the door. Preferably one that doesn't fall asleep during the night shift.

What is Project Perception?

A continuously operating, multi-model security system. In Microsoft's vision, it compiles the state of the environment from signals and logs, assesses risk, and then recommends or implements protective measures under human supervision.

Red, blue and green team

Red agents search for attack routes, blue agents investigate what constitutes a real threat, and green agents perform repairs and reinforcements. It's like a digital fire department, where one team looks for smoke, another determines if something is really burning, and the third already brings the hose.

What can MAI‑Cyber‑1‑Flash do?

The specialized model is designed to handle software vulnerabilities. According to Microsoft, the MDASH agent system achieved 96% in the CyberGym measurement, 12 points more than Mythos, while bringing nearly 50% cost savings compared to the current configuration.

When will it arrive?

Project Perception will begin public preview on August 3, 2026. This is primarily an enterprise security system, not a new Windows button that will politely leave all viruses with a click.

Why does the system still need a person?

A lot of strange but harmless events appear in the security logs. An automated system must distinguish a real attack from a misconfiguration, testing, or an unusual yet legitimate user action. If the system were to automatically shut down at every suspicious signal, the protection itself could cause a serious outage.

This is why privilege restrictions, logging, and human approval are important. AI can accelerate search and prioritize threats, but the business implications must also be understood before a major patch or network intervention.

What can this mean for smaller companies?

Project Perception is a corporate system, but the trend can also be seen in smaller services: faster alerts, more comprehensible summaries and repair suggestions. This does not change the basics. Updates, backups, multi-factor authentication, and appropriate permissions still can't be triggered by a flashy AI button.

After the news, come the frivolously serious part

Red, blue and green team in an office

The red team hacks the test system. The blue team spots it immediately. The green team will fix it in five minutes. All three sit back in satisfaction before someone asks, "And the printer on the third floor?"

It will be quiet. The printer has lived by its own rules since 2009, it only accepts paper on Tuesdays, and with the message "PC LOAD LETTER" it has already broken more security experts than any ransomware. Microsoft AI finally determines: there was no attack, the printer is simply like that.

Official source: Microsoft - Rethinking security for the age of AI. The benchmark results come from Microsoft's own announcement.