The article introduces Project Perception, a new agentic security system designed for the age of AI. It argues that traditional security approaches cannot keep up with AI-powered attacks, and proposes a system that continuously perceives, reasons, and acts at machine speed using specialized agents.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
When attackers operate at machine speed, traditional security architectures can no longer keep up. Microsoft introduces Project Perception, a closed-loop defense system composed of red, blue, and green agents, aiming to fight AI with AI.
Hayete Gallot, Corporate Vice President of Microsoft Security, stated in an official blog that autonomous systems can now reason, adapt, and operate continuously, while attack costs are decreasing, allowing attackers to generate vulnerabilities faster and launch attacks at greater scale. Traditional human-operated security methods cannot keep up with a world of AI, agents, and machine-speed attacks.
Therefore, Microsoft believes that the defining characteristic of next-generation security systems is no longer generating more alerts, but the ability to continuously sense, reason, and act. This directly led to the creation of Project Perception.
Project Perception is described as an 'agentic security system' that integrates signals, context, models, and specialized agents into a continuously learning defense system. Its core consists of three types of agents: Red Team agents identify potential intrusion paths before attackers exploit them; Blue Team agents investigate, reason about context, and determine meaningful risks; Green Team agents take corrective actions and strengthen defenses. The three work together in a closed loop, continuously discovering, assessing, and improving the security posture.
The underlying system is a new 'Cyber Stack,' starting from the signal and sensor layer, moving up through the Security Context layer, model layer, orchestration layer (Harness), agent layer, and execution layer. Microsoft particularly emphasizes the role of Security Context: it transforms Microsoft's visibility, threat intelligence, and security expertise into a near-real-time representation of assets, identities, relationships, risks, and activities that agents can directly use, thereby improving reasoning accuracy and reducing computational costs.
Project Perception employs a multi-model architecture, selecting the most suitable model for each security task, balancing quality, reliability, latency, and cost. Microsoft states that no single model is optimal for all security tasks.
The first application scenario is software vulnerability management. Microsoft integrated the MAI-Cyber-1-Flash model into MDASH (Multi-model Agentic team for Software Vulnerability). According to Microsoft, this configuration achieved 96% on the CyberGym benchmark, 12 percentage points higher than Mythos, while saving nearly 50% in costs compared to the current MDASH configuration. Microsoft plans to extend MAI-Cyber-1-Flash to more security workflows in the future.
Microsoft claims that the effectiveness of Project Perception depends on its visibility, actionable capabilities, team experience, and available models. Microsoft has broad visibility across identity, endpoints, applications, data, cloud, and AI systems, and can help customers take action across these environments. Additionally, decades of security research, threat intelligence, and real-world operational experience shape the system's reasoning and response.
Project Perception will enter public preview on August 3, 2026.
All information in this article comes from Microsoft's official blog, a first-party source. All performance data (96% benchmark score, +12 points, 50% cost savings) are self-reported by Microsoft and have not been independently verified. System architecture descriptions and preview timeline come from the same source.
Project Perception is Microsoft's security system designed for the AI era, operating through a closed loop of three types of agents, and for the first time discloses performance data under a multi-model strategy. However, all data are solely claimed by Microsoft, and actual effectiveness awaits third-party verification.
Primary report
Primary source