Project Perception: Microsoft created an AI that defends your systems by itself and almost never makes mistakes

The arrival of generative artificial intelligence is fundamentally changing both the offensive and defensive aspects of cyber security. AI allows attackers to find vulnerabilities faster, automate attacks and campaigns on a much larger scale than before – and traditional security tools, developed for environments where key decisions were made by humans, are having a hard time keeping up.

In response, Microsoft developed Project Perception – a platform that combines security signals, contextual data, AI models and specialized agents that can identify risks in real time, prioritize and take appropriate protection measures.

Three types of agents, three roles

Srž system čs three types of AI agents that work together. Red Team agents discover possible attack scenarios and weak points before attackers can exploit them. Blue Team agents analyze the security context and assess the actual level of risk. Green Team agents implement corrective measures and improve system protection. Working together, these three layers continuously cover the entire cycle of detection and response to threats.

Multi-model approach and concrete results

One of the key features of Project Perception is a multi-model approach – instead of one AI model, the system selects for each task the one that provides the best ratio of accuracy, speed, reliability and cost. The first example comes from software vulnerability management, where Microsoft integrated the MAI-Cyber-1-Flash model developed for specific security scenarios into the MDASH system. The new configuration achieved 96 percent success on the CyberGym benchmark – 12 percentage points higher than the Mythos model, with almost 50 percent lower operating costs compared to the current production configuration.

The platform draws on Microsoft’s experience in managing identities, devices, applications, data and AI systems, and is available through public preview from August 3, 2026.