Microsoft has unveiled its first custom-built cybersecurity AI model, MAI-Cyber-1-Flash, along with a new agentic defense platform called Project Perception. The company says these tools are designed to help enterprises find and fix software vulnerabilities more efficiently.
MAI-Cyber-1-Flash is a compact model developed by Microsoft's AI division. It scores 96% on CyberGym, a benchmark that tests AI reasoning over codebases, beating larger models like Gemini and GPT while cutting costs by about half compared to Microsoft's current system. According to the company, the model handles up to 90% of security tasks on its own, while difficult problems are escalated to OpenAI's GPT-5.4.
Project Perception coordinates three types of AI agents: red teams that hunt for vulnerabilities, blue teams that investigate risks, and green teams that remediate threats. It will enter public preview on August 3.
Microsoft AI CEO Mustafa Suleyman emphasized that the system's performance comes from the entire setup, not any single model. He also highlighted that token costs, not model quality, are the main barrier to enterprise AI adoption. The company processes over 100 trillion security signals daily, giving it a data advantage that Suleyman called a "moat."
Microsoft says access to the model will be tightly controlled to prevent misuse, with users screened for good intent and technical competence. The rollout will start with a small number of users and gradually expand.