Microsoft Unveils Project Perception, an Agentic Security System for the AI Era
Project Perception pairs red, blue, and green AI agent teams with a new MAI-Cyber-1-Flash model that scores 96% on the CyberGym benchmark.
Microsoft says the physics of cybersecurity are changing: autonomous systems can now reason, adapt, and operate continuously, while the cost of mounting an attack keeps falling. Attackers can generate exploits faster and scale campaigns further, and Microsoft argues that security approaches built for human-paced attackers cannot keep up with machine-speed ones. Its answer is Project Perception, which Microsoft describes as a new agentic security system built for the realities of AI, turning signals into real-time protections by using AI to defend against AI.
Three teams of agents
Project Perception coordinates three classes of specialized agents that work as a closed loop. Each group has a distinct job. The red team agents hunt for weaknesses an attacker could exploit, working proactively instead of waiting for an incident. The blue team agents then take that context and reason through it to judge which weaknesses actually matter. Finally, the green team agents step in to make fixes and harden the environment against the risks the other two groups surfaced. Microsoft says the result is a system that keeps discovering, judging, and improving an organization's defenses on an ongoing basis.
A cyber stack built from the ground up
Microsoft frames Project Perception as more than agents bolted onto existing workflows. It describes a full stack: signals and sensors that provide awareness across the digital estate, a security context layer that turns those signals into token-efficient understanding, models that provide intelligence and reasoning, a harness that coordinates models and agents, agents that apply that intelligence, and actuators that translate decisions into protection.
A multi-model approach, tested on vulnerability management
Microsoft says no single model is optimal for every security task, so Project Perception uses a multi-model architecture that mixes frontier and specialized cyber models. The first scenario Microsoft names is software vulnerability management: bringing its new MAI-Cyber-1-Flash model inside MDASH, described as Microsoft's software vulnerability multi-model team of agents. Microsoft reports that MDASH with MAI-Cyber-1-Flash delivers 96% on CyberGym, which it calls an industry-leading benchmark, 12 points above a system called Mythos, while cutting almost 50% of cost compared with the current MDASH configuration on the market today. Microsoft says it will next apply MAI-Cyber-1-Flash to more security workflows beyond vulnerability management.
Rollout
Microsoft is bringing Project Perception to customers worldwide, and says it enters public preview on August 3.
Key facts
- Project Perception is Microsoft's new agentic security system, combining red team, blue team, and green team AI agents in a closed loop.
- Its cyber stack has six layers: signals and sensors, security context, models, a harness, agents, and actuators.
- The first deployment pairs Microsoft's new MAI-Cyber-1-Flash model with MDASH, its software vulnerability team of agents.
- Microsoft reports MDASH with MAI-Cyber-1-Flash scores 96% on the CyberGym benchmark, 12 points above Mythos, at almost 50% lower cost than the current MDASH configuration.
- Project Perception enters public preview on August 3.
Check your understanding
What do the red team, blue team, and green team agents each do inside Project Perception? What score did MDASH with MAI-Cyber-1-Flash reportedly reach on the CyberGym benchmark, and how does that compare with Mythos?