Microsoft trained its first model built only for security. Not a general chatbot bolted onto a SOC — a cyber-specialized fine-tune of MAI-Code-1-Flash, sparse MoE, 137B total parameters but just 5B active, 256k context. Small and cheap on purpose.
What it actually does
MAI-Cyber-1-Flash lives inside MDASH, Microsoft’s multi-model vulnerability-scanning harness that discovers, validates, and patches bugs in real codebases. Swapping it in for 80% of the old models pushed MDASH from 88.4% to 95.95% on CyberGym — 1,500+ tasks across 188 projects. Pair it with GPT-5.4 for the hardest 10% and it lands 96%, at roughly half the cost. It also autonomously surfaced 16 Windows CVEs, including several remote-code-execution bugs.
The API angle
Delivered through Project Perception, Microsoft’s agentic defense platform, with public preview on August 3, 2026 — expect Azure/Foundry API access. Point it at a repo to simulate attacks, triage threats, and auto-fix vulns. The bet: let a tiny model do 90% of the grunt work and only escalate the nasty cases. Defense automation, timed for the AI-agent-escape headlines.
You Might Also Like
- Microsoft Mdash Scores 88 45 on Cybergym Beating Anthropic Mythos and Openai gpt 5 5
- Microsoft mai Code 1 Flash is Live in Github Copilot 60 Fewer Tokens Than Comparable Coding Models
- Microsoft Aion 1 0 Puts a 14b Agent Model Inside Windows no Cloud Required
- Microsoft mai Models mai Transcribe 1 mai Voice 1 mai Image 2 are Live Redmonds ai Independence Starts now
- Microsoft Agent Governance Toolkit Scores 10 10 on Owasp Agentic Risks at 0 1ms per Check

Leave a comment