Microsoft says it has introduced its first cybersecurity-specific model within MDASH, its multi-model system for identifying and remediating software vulnerabilities. According to the company, MDASH incorporates MAI-Cyber-1-Flash alongside other models to support automated vulnerability discovery. Microsoft states that, in testing on CyberGym, the updated setup reaches 95.95% performance. Microsoft also compares the cost of configuring the new model against what it describes as its current best MDASH combination, which uses GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex. The company claims the new configuration costs about half as much while maintaining or improving results.
Decrypt additionally reports Microsoft’s claim that the system enables more than 100 AI agents to find software flaws and that the new approach performs favorably versus other models in the same kind of cybersecurity evaluation. The reporting is based on Microsoft’s figures and does not include independent verification in the provided excerpts.
