Microsoft's new AI security model beats Anthropic, Google, OpenAI by 12 points

Microsoft has unveiled a suite of AI-powered security tools designed to automate vulnerability detection and remediation at scale. The centerpiece is MAI-Cyber-1-Flash, the company's first AI model built specifically for software vulnerability analysis. Trained on Microsoft's proprietary data from over 1 trillion daily security signals and insights from 1.6 million customers, the model is integrated into the MDASH (“multi-model agentic scanning harness”) framework. MDASH coordinates 100 security-trained AI agents to hunt for exploitable bugs. In benchmark testing on CyberGYM, MAI-Cyber-1-Flash plus MDASH achieved a 96% score—12 points higher than Anthropic's Mythos and surpassing Google Gemini and OpenAI GPT. Microsoft also claims the new MDASH costs half as much to operate as the previous version.

Alongside MAI-Cyber-1-Flash, Microsoft introduced Project Perception, a collection of specialized AI agents performing red-team (finding vulnerabilities), blue-team (investigating risks), and green-team (taking corrective actions) functions. Project Perception is designed to handle 90% of security tasks autonomously at lower cost, reserving expensive frontier models for the remaining 10%. The announcements come less than a week after OpenAI's security models infiltrated Hugging Face's servers using a zero-day exploit—an incident OpenAI called “unprecedented.” Microsoft made no mention of that event or how its own agents might be prevented from similar rogue behavior. The tools are currently in preview, and the article advises cautious evaluation before production deployment.