Microsoft has unveiled Project Perception, a new agentic security system built to help defend against AI-powered cyberattacks, alongside its first-ever custom-built cybersecurity model.
The announcement came during a security-focused event in San Francisco, marking Microsoft’s biggest push into AI-driven cybersecurity since a leadership shake-up earlier this year.
Microsoft Unveils Project Perception: New Cybersecurity Model
Microsoft’s blog post announcing the launch of Project Perception explained that the physics of cybersecurity are changing as autonomous systems grow capable of reasoning and operating continuously, while the cost of launching an attack keeps falling.
According to Microsoft, the old approaches built for human attackers can no longer keep pace with a world of AI agents operating at machine speed, prompting the company to build what it calls a new Cyber Stack.
At the center of that stack is MAI-Cyber-1-Flash, Microsoft’s first AI model built specifically for cybersecurity work. The company said the model went through review by its internal AI Red Team, adversarial testing, and evaluation from an outside third party before release.
Microsoft AI CEO Mustafa Suleyman said during the event that the company is delivering world-leading performance at half the cost of comparable systems.
Project Perception: Fighting AI with AI
Project Perception itself functions as an agentic security system, coordinating three distinct groups of AI agents working together. Red team agents search for paths an attacker could realistically exploit, blue team agents determine which of those risks actually matter, and green team agents carry out the fixes needed to close the gaps.
Microsoft’s executive vice president of Microsoft Security, Hayete Gallot, described the system as one that brings together signals, context, models, and specialized agents into a continuously learning system of defense, capable of reasoning and acting at machine speed while keeping human oversight in place.
Microsoft says Project Perception will enter public preview on August 3, with plans to expand the system with additional specialized security agents over time. The company positioned the tool as a way for enterprise customers to access continuous AI-powered security monitoring through a single platform.
This aims to help forego the need to build custom agents in-house or contract separately with other AI providers for top-tier cybersecurity models.
Project Perception Combines Two Previous Models
MAI-Cyber-1-Flash runs inside MDASH, Microsoft’s existing multi-agent harness for finding and fixing software vulnerabilities and is built on top of the company’s internally developed MAI-Thinking-1 reasoning model. For roughly 90 percent of tasks, Microsoft relies on MAI-Cyber-1-Flash alone, reserving OpenAI’s GPT-5.4 for the remaining 10 percent of cases the company considers exceptionally difficult.
According to Microsoft’s own benchmark testing, this combined configuration scored 96 percent on CyberGym, an industry benchmark measuring how well AI systems reason through large codebases to find real vulnerabilities.
The company places its model 12 points above Anthropic’s Mythos 5 and ahead of OpenAI’s GPT-5.5 Cyber and Google’s Gemini 3.5 Flash Cyber.
Microsoft said that same configuration also delivers close to 50 percent in cost savings compared to its current MDASH setup already in market. They also framed the combination of strong benchmark results and lower operating cost as the core pitch behind Project Perception’s broader rollout.
