Microsoft is expanding its artificial intelligence portfolio into cybersecurity with the launch of a new agentic cybersecurity platform alongside its first AI model built specifically to find software vulnerabilities.
Announced Monday, Project Perception and MAI-Cyber-1-Flash are designed to help security teams identify vulnerabilities, prioritize risks, and respond to cyber threats faster by using AI agents that work together across an organization’s environment. The announcement comes as cybercriminals increasingly use artificial intelligence to automate attacks, forcing security teams to rethink how they protect enterprise networks.
The launch also highlights an escalating race among major AI companies to build defensive AI systems before cybercriminals gain a lasting advantage from the technology. As AI becomes more capable of discovering vulnerabilities and launching attacks, technology companies are racing to build defensive AI systems that can keep up.
Microsoft introduces an AI-powered cyber defense platform
Project Perception is designed to automate security work by coordinating specialized agents across three teams. Red team agents simulate attacks and search for weaknesses, blue team agents investigate the findings and assess their severity, and green team agents recommend or apply remediations.
Microsoft said this structure creates a continuous cycle in which vulnerabilities can be identified, prioritized, and addressed without adding more alerts to analysts’ queues. The platform can also route individual tasks to different AI models based on factors including accuracy, speed, reliability, and cost.
The company has positioned the platform as a way to compress work that might otherwise require several security specialists and hours of manual investigation. Project Perception is scheduled to enter public preview on August 3.
New cybersecurity model focuses on finding software vulnerabilities
Microsoft also introduced MAI-Cyber-1-Flash, its first AI model developed specifically for cybersecurity applications. The model works with MDASH, Microsoft’s software vulnerability management framework, to search complex codebases for difficult-to-detect flaws.
Microsoft AI CEO Mustafa Suleyman said a configuration combining MAI-Cyber-1-Flash with GPT-5.4 scored 96% on the CyberGym benchmark. According to Suleyman, it outperformed configurations using competing models from Anthropic, Google, and OpenAI while costing nearly half as much as Microsoft’s previous setup.
Those results are Microsoft’s own benchmark claims and would benefit from independent testing. The combined configuration also makes it difficult to determine from the score alone how MAI-Cyber-1-Flash would perform as a standalone model.
AI is changing the cybersecurity arms race
Microsoft’s announcement comes as leading AI developers test increasingly capable systems for both offensive and defensive cybersecurity work. The New York Times reported this week that two OpenAI systems found and exploited vulnerabilities in a widely used internet library during controlled internal testing.
The experiment illustrates the tension facing AI developers: The same capabilities that can help defenders locate vulnerabilities may also allow attackers to discover and exploit them more quickly.
Microsoft is entering a growing market that includes cybersecurity initiatives from Anthropic, OpenAI, Google, and specialized security vendors. These offerings vary considerably, ranging from commercial security platforms to research systems and internal experiments.
As AI transforms both offensive and defensive cybersecurity, Microsoft’s latest announcement signals where the industry is headed. Instead of simply helping analysts work faster, the next generation of cybersecurity will rely less on isolated security tools and more on coordinated AI agents that continuously identify, evaluate, and mitigate threats alongside human analysts.
Read next: Microsoft is also expanding its in-house AI portfolio across workplace applications. Learn how its MAI models could replace OpenAI technology in Excel and Outlook while reducing costs.