Microsoft Says MDASH Beats Claude Mythos and GPT-5.6 Sol in Cybersecurity Test
Microsoft introduced MAI-Cyber-1-Flash, its first dedicated cybersecurity model, and paired it with the MDASH vulnerability-hunting system. Microsoft says the combined setup scored 95.95% on CyberGym, outperforming reported scores for GPT-5.5 Cyber, Anthropic’s Mythos 5, GPT-5.6 Sol, and Gemini 3.5 Flash Cyber, while costing about 50% less than Microsoft’s prior best MDASH configuration. MAI-Cyber-1-Flash is not used alone: it handles up to 90% of the work, while MDASH sends the hardest cases to GPT-5.4. MDASH is a multi-agent harness with over 100 specialized agents that audit code, debate findings, remove duplicates, and build proof-of-concept exploits to confirm real vulnerabilities. Microsoft is positioning this as a shift toward cheaper, more scalable bug discovery, arguing that validation and proof are now the main bottlenecks. MDASH is entering private preview in Microsoft Security Exposure Management, with repository scanning, ranked findings, and proposed fixes available through Defender tools.
