Microsoft Corporation Launches Project Perception and MAI-Cyber-1-Flash Security AI

Microsoft Corporation launched Project Perception and a specialized AI model, MAI-Cyber-1-Flash, to shift enterprise security from reactive investigation to continuous risk reduction. Project Perception employs a Cyber Stack architecture utilizing three types of AI agents: red agents to identify attack paths, blue agents to investigate risks, and green agents to execute corrective actions.

The system integrates the compact MAI-Cyber-1-Flash model—developed by the internal FORGE team led by Taesoo Kim—with OpenAI's GPT-5.4 via a harness called MDASH. Microsoft Corporation claims this combination reduces operational costs by approximately 50% while outperforming competitors such as Anthropic's Mythos 5 and Google's 3.5 Flash Cyber on the CyberGym benchmark, where the new model scored 96%.

Executive Vice President of Security Hayete Gallot stated the system is necessary because attackers are leveraging AI to scale campaigns and reduce the cost of offense. Project Perception enters public preview on August 3 via Microsoft Defender using a consumption-based pricing model, while the MAI-Cyber-1-Flash model will be available through Azure AI Foundry.