Microsoft MAI-Cyber-1-Flash, its first AI model built specifically for cybersecurity. The smaller, specialized model sits inside MDASH, Microsoft's multi-agent system for finding and fixing software vulnerabilities. Microsoft says the combination scores 95.95 percent on the CyberGym benchmark, about 12 points above Anthropic's Mythos 5, while cutting vulnerability-detection costs by nearly 50 percent. MAI-Cyber-1-Flash handles roughly 90 percent of the work, including detection, patch generation and verification. Only the hardest cases are passed to OpenAI's GPT-5.4. Alongside the model, Microsoft introduced Project Perception, an AI platform that continuously scans systems for risks. It uses three types of agents: red-team agents that hunt for weaknesses, blue-team agents that prioritise alerts, and green-team agents that propose or, with approval, apply fixes. Project Perception enters public preview on 3 August. MAI-Cyber-1-Flash will be available through Azure AI Foundry.
Microsoft MAI-Cyber-1-Flash, its first AI model built specifically for cybersecurity. The smaller, specialized model sits inside MDASH, Microsoft's multi-agent system for finding and fixing software vulnerabilities. Microsoft says the combination scores 95.95 percent on the CyberGym benchmark, about 12 points above Anthropic's Mythos 5, while cutting vulnerability-detection costs by nearly 50 percent. MAI-Cyber-1-Flash handles roughly 90 percent of the work, including detection, patch generation and verification. Only the hardest cases are passed to OpenAI's GPT-5.4. Alongside the model, Microsoft introduced Project Perception, an AI platform that continuously scans systems for risks. It uses three types of agents: red-team agents that hunt for weaknesses, blue-team agents that prioritise alerts, and green-team agents that propose or, with approval, apply fixes. Project Perception enters public preview on 3 August. MAI-Cyber-1-Flash will be available through Azure AI Foundry.