微软宣布推出MAI-Cyber-1-Flash模型，并将其集成到MDASH多智能体漏洞识别和修复平台中[1]。该系统在网络安全基准测试CyberGym上达到96%的性能，较Mythos提高12个百分点，并超越Gemini和GPT等竞品[1]。
在成本控制方面，新方案相比现有MDASH最优方案实现了50%的成本节省[1]。这一优势源于系统的智能设计——MDASH能够高效处理90%的任务，仅对最具挑战性的10%任务调用成本最高的模型[1]。MDASH平台包含由安全专家创建的100多个智能体，该模型已经过微软AI红队的严格评估和第三方独立评估[1]。
微软同时推出Perception智能安全系统，用于持续监控和修复威胁[1]。根据微软披露的数据，该公司每天观察到来自超过100万客户的超过100万亿个安全信号[1]。
Microsoft has announced the introduction of the MAI-Cyber-1-Flash model, integrated into MDASH, a multi-agent platform designed for vulnerability identification and remediation.[1] The system achieves 96% performance on the CyberGym cybersecurity benchmark, surpassing competing models including Mythos, Gemini, and GPT while reducing costs by 50% compared to the current optimal configuration for MDASH.[1]
The architecture of the new system demonstrates efficiency through task stratification, handling 90% of tasks with more cost-effective models while reserving the highest-cost model exclusively for the most challenging 10% of tasks.[1] MDASH comprises over 100 agents developed by security experts to identify and address threats.[1] Microsoft also unveiled Perception, an intelligent security system designed for continuous threat monitoring and remediation.[1]
The deployment comes as Microsoft processes security signals at scale, observing over one million trillion security signals daily across more than one million customers.[1] Both the MAI-Cyber-1-Flash model and the broader system architecture have undergone rigorous evaluation by Microsoft's AI red team as well as independent third-party assessment.[1]