Microsoft AI's first cybersecurity model: a sparse MoE (137B total / 5B active, 256K context) in the MAI-Thinking-1 lineage. Inside the MDASH multi-agent security harness it handles ~90% of vulnerability-identification work (GPT-5.4 takes the remainder); MDASH scores ~96% on CyberGym, claimed +12 points over Anthropic Mythos at half the cost. Proprietary.

MAI Cyber 1.1 Flash (October 27, 2026) replaces it as a selectable scan profile in MDASH. Microsoft reports performance comparable to 1 Flash at 18% lower cost per scan on its internal vulnerability-scanning suite, with no absolute scores; access stays limited to verified MDASH customers. Parameters are not restated for 1.1.

Model Details

Architecture MOE
Parameters 137B
Active params 5B
Context window 256,000
Base model mai-thinking-1

Variants

Name Parameters Notes
MAI-Cyber-1-Flash — Jul 27 2026; 137B total / 5B active MoE
MAI-Cyber-1.1-Flash — Oct 27 2026; comparable to 1 Flash, 18% lower cost per scan (internal MDASH suite)
agentsmoesecurity

Related