Microsoft Launches First Cyber Model and Agentic Security System

Finding software vulnerabilities is getting faster. Fixing them is still slow.

Microsoft has introduced MAI-Cyber-1-Flash, its cybersecurity-focused AI model, alongside Project Perception, an agentic system designed to connect vulnerability discovery, investigation and remediation.

The model works inside MDASH Microsoft’s -agent environment for finding and validating software flaws. MAI-Cyber-1-Flash is a code-focused model derived from the MAI-Thinking-1 family. Microsoft says it can handle up to 90% of cyber tasks while more difficult cases are passed to larger models such as GPT-5.4.

This structure is intended to reduce operating costs without sacrificing performance. Microsoft reports that the combined MDASH configuration scored 95.95% on CyberGym around 12 percentage points above Mythos while costing half as much as its previous setup. These are company-reported benchmark results.

Project Perception takes the system beyond vulnerability scanning. It coordinates three groups of agents:

Red-team agents look for attack paths.

Blue-team agents investigate the context. Decide which findings represent real risk.

Green-team agents support remediation by improving detection changing security settings or preparing code fixes.

Microsoft wants the system to create a loop between finding a problem and responding to it rather than producing another alert for security teams to review manually. Human approval remains important where changes could affect production systems.

The company says it processes than 100 trillion security signals each day and supports around 1.6 million security customers. MDASH also includes, than 100 specialized agents built around different models.

That scale gives Microsoft a source of operational security data but agentic remediation introduces new risks. A system that changes code or security controls could create problems if it acts on information. Microsoft says MDASH includes role-based access, tenant isolation, encryption, auditing and sandboxed execution.

Project Perception is expected to enter preview on August 3 2026 according to Microsoft’s official announcement.

The main test is not whether Microsoft’s AI can find vulnerabilities. It is whether the system can help companies fix them faster without creating security or operational problems.