Microsoft ships a cybersecurity model and an agent security system

Microsoft shipped MAI-Cyber-1-Flash on July 27, a cybersecurity model trained on its own exploit and remediation records.

Microsoft says the model runs inside MDASH at half the cost of leading models, per its announcement. That cost claim matters as much as the security claims for anyone comparing model spend.

Project Perception, the agent platform released alongside it, runs red-team, blue-team, and green-team agents in the same workflow. Red-team agents hunt for compromise paths, blue-team agents triage what they find, and green-team agents remediate.

On CyberGym, the model scored 96%, which Microsoft says beats Anthropic's Mythos by 12 points, per the announcement.

Public preview lands August 3 inside Defender. The model also becomes available the same day on Azure AI Foundry, per the announcement.

The benchmark caveat

A 12-point CyberGym gap over a named competitor deserves a second look before it moves a purchase decision. Microsoft's announcement doesn't say who ran the benchmark or whether the comparison used matched conditions.

What's clear is the structural bet. Microsoft packaged red, blue, and green agent roles as one runtime instead of shipping a model and leaving orchestration to security teams.