Microsoft reports its MDASH system achieved a 95.95% score on CyberGym, surpassing competitors while reducing costs by 50%.
Microsoft announced its MDASH vulnerability-hunting system scored 95.95% on the CyberGym benchmark, outperforming rivals including Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol. The result, self-reported by Microsoft, exceeds GPT-5.5 Cyber’s 85.6% and Gemini 3.5 Flash Cyber’s 83.2%.
CyberGym evaluates AI agents by reproducing 1,507 known vulnerabilities across 188 open-source projects. Microsoft’s MAI-Cyber-1-Flash model handles 90% of tasks, while MDASH routes the remaining 10% to GPT-5.4 for complex cases.
The company claims the new setup costs 50% less than its previous best MDASH configuration. The system is currently in private preview through Microsoft Defender, allowing teams to review findings and generate fixes.