Microsoft's new multi-model agentic security system tops leading benchmark (www.microsoft.com)

Microsoft has unveiled a groundbreaking advancement in AI-driven cybersecurity with its multi-model agentic security system, codename MDASH, which successfully identified 16 new vulnerabilities in Windows networking and authentication components, including four critical remote code execution flaws. MDASH utilizes over 100 specialized AI agents orchestrated across various models to ensure comprehensive bug discovery, boasting an impressive score of 88.45% on the CyberGym benchmark—the highest in the industry. This innovation signifies a shift from experimental research to practical, enterprise-scale defenses that could reshape how organizations handle vulnerability identification and mitigation. The architecture of MDASH represents a critical evolution in vulnerability assessment, employing diverse models tailored for different auditing roles, from discovery to validation. With its end-to-end pipeline, the system integrates plugins that allow expert input on complex vulnerabilities, ensuring broad adaptability and continual improvements as AI models evolve. By achieving zero false positives in a test with deliberately injected vulnerabilities, the system has shown its potential to match or exceed human researchers in effectiveness. This leap forward in AI-enabled security tools not only enhances Microsoft’s defense mechanisms but may set new standards for the entire cybersecurity industry.