Microsoft’s MDASH outperforms Claude Mythos, GPT-5.6 in cybersecurity test

Microsoft just dropped what might be the most compelling argument yet for why the future of AI security isn’t about building one really smart model. It’s about building a hundred of them and teaching them to work together. The company unveiled MDASH, short for Microsoft Security multi-model agentic scanning harness, a cybersecurity system that orchestrates over 100 specialized AI agents to hunt down and fix vulnerabilities across complex codebases. On the CyberGym benchmark, MDASH scored 88.45%, comfortably beating Anthropic’s Claude Mythos Preview at 83.1% and OpenAI’s GPT-5.5 at 81.8%. A swarm beats a single brain Rather than relying on a single large language model to do everything, Microsoft’s Autonomous Code Security team built an orchestrated system where each agent handles a specific slice of the vulnerability detection and remediation process. The approach already proved its worth during the May 2026 Patch Tuesday, where MDASH identified 16 new Windows vulnerabilities. Four of those were critical remote code execution flaws. Microsoft’s team developed the system drawing on experience from DARPA’s AI Cyber Challenge. The system is integrated with Microsoft Defender and GitHu...