Microsoft AI released MAI-Cyber-1-Flash on July 27 and embedded it into MDASH, its multi-agent vulnerability identification and remediation harness. The headline number is strong: MDASH with MAI-Cyber-1-Flash reached 95.95%, rounded to 96%, on CyberGym.
The split matters
Microsoft says the new model is designed to handle up to 90% of tasks efficiently, while MDASH routes the hardest 10% to larger models, currently GPT-5.4. The company claims this delivers 50% cost savings against its best current MDASH setup.
That cost angle is the real story. Cybersecurity teams cannot treat AI scanning as a rare demo. Codebases, dependencies and attacks move continuously, so the unit economics of repeated scans matter.
What was verified
Microsoft's chart puts the combined system at 95.95%, with other models shown between 83.2% and 85.6%. Help Net Security independently reported Microsoft's claim that the system outperformed Mythos, Gemini and GPT, and that the model underwent AI Red Team review, adversarial testing and third-party assessment.
Microsoft also links the release to its earlier MDASH work, where the older system scored 88.45% on a CyberGym set of 1,507 real vulnerabilities, found 21 of 21 planted vulnerabilities with no false positives in a private driver, and reported 96% recall on clfs.sys MSRC cases and 100% on tcpip.sys.
Why enterprises care
Project Perception is the second part of the release. It uses MDASH as an agentic security system and is scheduled for public preview on August 3, 2026. The control layer is specific: role-based controls, tenant isolation, encryption, auditability and sandboxed execution with no internet access.
The next useful checks are not more launch claims. They are customer cases, third-party retests, false-positive rates and whether the system can move from finding flaws to validating fixes inside enterprise workflows.
Sources: Microsoft AI official release, Microsoft Security Blog, Help Net Security, CocoLoop; verified MAI-Cyber-1-Flash/MDASH release scope, 95.95%/96% CyberGym score, 12-point lead over Mythos, 50% cost saving, 90%/10% task routing, 100+ agents, 100 trillion daily security signals, 1.6 million customers and Project Perception preview plan.