Microsoft MDASH tops CyberGym with 100+ agents

MDASH marks a definitive shift from general-purpose reasoning to orchestrated multi-agent ensembles for high-stakes security engineering. By implementing a formal "debate and prove" architecture, Microsoft has successfully moved autonomous security tools from simple pattern matching to verifiable vulnerability research.

• –Employs a tiered pipeline of auditor, debater, and prover agents to eliminate false positives and confirm exploitability.
• –Leverages a hybrid model strategy, combining the reasoning depth of frontier models with the speed of task-specific distilled models.
• –Demonstrated real-world efficacy by identifying 16 new Windows vulnerabilities, including 4 critical kernel remote code execution flaws.
• –The 88.45% CyberGym score establishes a new state-of-the-art for offensive security agents in open-source and enterprise codebases.