Microsoft's First Cybersecurity Model: 90% of Tasks Handled by In-House Small Model, Costs Reduced by Half

According to monitoring by Beating, Microsoft has released its first self-developed cybersecurity model, MAI-Cyber-1-Flash, and integrated it into the vulnerability scanning system MDASH. This model is a cybersecurity fine-tuned version of MAI-Code-1-Flash, with a total of 137 billion parameters, activating 5 billion parameters per instance, and supporting a 256K context window. MDASH enables it to handle up to 90% of tasks related to vulnerability detection, validation, classification, and remediation. The most challenging 10% of tasks are delegated to GPT-5.4. Microsoft claims that this combination is nearly 50% cheaper compared to existing solutions like GPT-5.4, GPT-5.4 mini, and GPT-5.3 Codex. In Microsoft's CyberGym tests, the "MAI-Cyber-1-Flash + GPT-5.4" scored 95.95%, while GPT-5.5 Cyber scored 85.6% and Anthropic Mythos 5 scored 83.8%. It should be noted that this comparison involves complete model and agent combinations, and does not imply that MAI-Cyber-1-Flash alone outperformed these models. Additionally, it will be integrated into the security agent product Project Perception. The red, blue, and green agents will respectively handle tasks such as vulnerability detection, risk assessment, and defense reinforcement. Project Perception is set to launch public testing on August 3, but MAI-Cyber-1-Flash is currently only available for private testing to MDASH clients who pass the review process. [Original Link]