Microsoft Unveils Cyber AI Model, Agentic Security Platform

Cybersecurity vendors are racing to automate defense as attackers use generative AI to discover and exploit software flaws faster. Microsoft’s Project Perception moves beyond a security copilot by coordinating Red, Blue and Green agent teams to probe code, rank threats and prepare or deploy fixes, while requiring human approval for high-impact actions. The platform builds on MDASH, Microsoft’s multi-model vulnerability scanning harness, which uses more than 100 specialized agents and in May helped uncover 16 previously unknown Windows flaws.

Microsoft unveiled MAI-Cyber-1-Flash and Project Perception in San Francisco on July 27, 2026. The in-house cyber model has 5 billion active parameters and handles about 90% of MDASH’s workload, routing the hardest 10% to OpenAI’s GPT-5.4. Microsoft said the combined system scored 95.95% across 1,507 CyberGym tasks, versus roughly 83% for rival models, while cutting operating costs by about 50% from its previous top configuration. Perception is due to enter public preview in Microsoft Defender on Aug. 3, when the model will also become available through Azure AI Foundry to vetted customers.