Release2026-07-27

Microsoft introduced MAI-Cyber-1-Flash, a compact model trained to find security flaws in large codebases, and put it inside MDASH, its multi-agent system for identifying and fixing vulnerabilities. The smaller model handles most of the work and hands the hardest cases to a larger one, which Microsoft says halves the cost of its previous setup while scoring higher on the CyberGym test. Microsoft also launched Perception, a set of security agents that will use the model for further workflows.

What changed

MDASH's best configuration combined GPT-5.4, GPT-5.4 mini and GPT-5.3 codex, scoring around 84% on CyberGym.

What it unlocks

Running continuous automated scanning of large codebases for security flaws at roughly half the previous processing cost.

  • 96% on CyberGym (+12 points above Mythos)
  • 50% cost saving versus the current best MDASH configuration
  • handles up to 90% of tasks, with the largest model reserved for the remaining 10%
  • more than 100 trillion security signals daily
  • 100+ agents in MDASH

What you need to act on it

  • access to MDASH, Microsoft's multi-agent security system

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.