What was announced
Microsoft introduced Project Perception, an agentic security system that connects signals across identities, endpoints, applications, data, and clouds into shared security context. Specialized models and agents then discover, assess, and remediate risk. Public preview is scheduled for August 3, 2026.
Three agent roles
- Red-team agents search for potential compromise paths before attackers exploit them.
- Blue-team agents combine signals with organizational context to determine meaningful risk.
- Green-team agents apply approved corrections and strengthen defenses.
Microsoft emphasizes a multi-model architecture that chooses frontier or cyber-specialized models according to quality, reliability, latency, and cost rather than relying on one model.
Reported performance
The first scenario is software vulnerability management. Microsoft says MDASH with MAI-Cyber-1-Flash scored 96% on CyberGym, 12 percentage points above its Mythos comparison, while cutting costs by nearly 50% versus the current market configuration.
Limitations
These performance and cost figures are vendor-reported results from Microsoft's selected benchmark and configuration. Real deployments must separately test permissions, false positives, remediation side effects, approval latency, and integration with existing security controls. Automated remediation requires human approval and a recovery path.