Microsoft has announced Project Perception, an AI-powered agentic security system designed to help organisations defend against AI-driven cyber threats. The platform combines specialised AI agents, security context, AI models and automated workflows to turn security signals into real-time protections while keeping humans in control of critical decisions.
Project Perception: AI agents for continuous security
Project Perception uses a coordinated workforce of specialised AI agents that analyse an organisation’s security environment and work together to improve its overall security posture.
The system includes three categories of agents:
- Red Team agents simulate attacker behaviour to identify potential attack paths and vulnerabilities before they can be exploited.
- Blue Team agents investigate security incidents, analyse contextual information and determine which risks require attention.
- Green Team agents take corrective actions to remediate identified issues and strengthen an organisation’s security posture.
According to Microsoft, these agents operate as a continuous closed-loop system that discovers, evaluates and responds to security risks instead of simply generating alerts.
Cyber Stack combines AI models, context and automation
Project Perception is built on Microsoft’s new Cyber Stack, which brings together signals, security context, AI models, orchestration and automated actions into a continuously learning security system.
The Cyber Stack includes:
- Signals and sensors that collect security data across identities, endpoints, applications, data, cloud environments and AI systems.
- Security context, which connects assets, identities, relationships, risks and activities to provide AI agents with a shared, token-efficient understanding of an organisation’s environment without repeatedly processing raw security data.
- A multi-model AI architecture that combines frontier AI models with specialised cybersecurity models, selecting the most suitable model for each task based on factors including quality, reliability, latency and cost.
- An orchestration harness that coordinates AI models and specialised agents across security workflows.
- Actuators that translate AI-generated decisions into actions across Microsoft Security products, enabling organisations to reduce risks rather than only identify them.
Microsoft says this architecture enables Project Perception to continuously understand risk, adapt to changing environments and automate security workflows more efficiently.
MAI-Cyber-1-Flash debuts for vulnerability management
As part of its multi-model strategy, Microsoft introduced MAI-Cyber-1-Flash, a specialised cybersecurity AI model that is initially being deployed within MDASH, Microsoft’s software vulnerability management system.
According to Microsoft, MDASH powered by MAI-Cyber-1-Flash achieved 96% on the CyberGym benchmark, which the company says is 12 percentage points higher than Mythos. Microsoft also says the configuration delivers nearly 50% lower operating costs than the current MDASH configuration and plans to extend MAI-Cyber-1-Flash to additional security workflows beyond software vulnerability management.
Availability and pricing
Microsoft says Project Perception will enter public preview on August 3. It will initially be available through Microsoft Defender, with support for additional Microsoft Security products planned in the future.
Project Perception follows a consumption-based, pay-as-you-go pricing model. Usage is measured in Security Compute Units (SCUs), with different AI agents consuming varying numbers of SCUs depending on the complexity of the tasks they perform.