Microsoft Says New Cybersecurity AI Model Helps MDASH Score 95.95% at Half the Cost - The Hacker News

What the report says

The Hacker News reports that Microsoft has introduced MAI-Cyber-1-Flash, its first cybersecurity-focused AI model for use inside MDASH, the company’s multi-model system for vulnerability identification and remediation. Microsoft says the new setup, which pairs MAI-Cyber-1-Flash with GPT-5.4, achieved a 95.95% score on CyberGym while costing 50% less than its current best MDASH configuration using GPT-5.4, GPT-5.4 mini and GPT-5.3 Codex.

According to the report, MAI-Cyber-1-Flash is intended to take on as much as 90% of MDASH work, while GPT-5.4 is used for the most difficult remaining tasks. Access is restricted to approved MDASH customers through an Azure AI Foundry private preview, and the model is not being offered as a standalone public model or general-purpose API. Microsoft’s model card describes it as a sparse mixture-of-experts transformer with 137 billion total parameters, five billion active parameters and a 256,000-token context window.

The Hacker News notes important limits around the benchmark claims. The 95.95% figure applies to the MDASH system using both models, not to MAI-Cyber-1-Flash alone. CyberGym Level 1 evaluates whether an agent can reproduce a known vulnerability and generate a working proof of concept from provided code and a description; it does not test blind vulnerability discovery or patch correctness. The result reportedly was not listed on CyberGym’s public leaderboard when checked on July 28, 2026.

Microsoft says the model was tested in an isolated environment and warns that generated code or text may need review before consequential use. The launch is tied to Project Perception, Microsoft’s broader effort to coordinate defensive security agents, which is scheduled for public preview on August 3.