Microsoft Ships Its First Cybersecurity Model and a Red/Blue/Green Agent Platform — 96% on CyberGym at Half the Cost

Microsoft AI (corroborated by TechCrunch, SiliconANGLE, VentureBeat and Axios)·high signal

Microsoft launched MAI-Cyber-1-Flash and Project Perception on July 27 at a small San Francisco event. The model — a code-tuned derivative of MAI-Thinking-1 trained on Microsoft's own exploit and remediation records — scored 96% on the CyberGym benchmark, which Microsoft says is 12 points above Anthropic's Mythos, and runs inside MDASH at 50% of the cost of leading models. Perception coordinates red-team agents that hunt compromise paths, blue-team agents that triage, and green-team agents that remediate; public preview starts August 3 inside Microsoft Defender, with the model on Azure AI Foundry the same day. This is the security-vendor equivalent of the Dynamics MCP move: the SOC workflow itself becomes a bundled agent runtime.