Microsoft Launches MAI-Cyber-1-Flash Inside MDASH
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Microsoft launched MAI-Cyber-1-Flash, its first cybersecurity-specific model, inside MDASH and routed it to handle up to 90% of tasks while reserving GPT-5.4 for the hardest 10%, per the company's model card.
- MDASH running MAI-Cyber-1-Flash alongside GPT-5.4 scored 95.95% on CyberGym Level 1, but the article stresses this is a system-level score, not a standalone result for the new model.
- CyberGym's public leaderboard did not list Microsoft's 95.95% when checked on July 28, 2026; Wiz's Atlas agent topped the board at 90.9% in a July 27 entry, while Microsoft's earlier MDASH submission sat at 88.4%.
- MAI-Cyber-1-Flash is a sparse mixture-of-experts transformer with 137 billion total parameters, 5 billion active, and a 256,000-token context window, fine-tuned from MAI-Code-1-Flash off a MAI-Thinking-1 mid-training checkpoint.
- Microsoft claims the configuration costs 50% less than its current best MDASH mix, but disclosed no token use, call volume, latency, task mix, or compute allocation behind that comparison.
- MAI-Cyber-1-Flash scored zero across the kernel, userspace, and browser categories of ExploitGym and posted 0.314 on CVEBench, 0.553 on CyberSecEval4 threat intelligence, 0.33 on malware analysis, and 0.651 on CRSBench at POV=1200.
- Project Perception, Microsoft's broader system for coordinating defensive security agents, is scheduled to enter public preview on August 3, with MAI-Cyber-1-Flash inside MDASH its first announced scenario.
Why it matters: Microsoft's headline 95.95% figure is a system-level MDASH score rather than a standalone benchmark for MAI-Cyber-1-Flash, and the model posted zeros on ExploitGym's exploit-generation categories — meaning buyers evaluating the private preview are buying a routing claim and a cost ratio, not yet a reproducible, independently published result.
