Microsoft and Wiz Deploy Multi-Model AI Agents to Hunt Zero-Days
Both companies report breakthrough detection rates on CyberGym benchmark while cutting costs by half through model orchestration.
Microsoft and cloud security firm Wiz have unveiled competing multi-model AI systems that outperform single-model approaches at finding software vulnerabilities, marking a shift toward agentic security orchestration over raw model scale.
Microsoft's MDASH (Multi-Model Defense Agentic Security Harness) and Wiz's Project Atlas both exceeded 88% success rates on the CyberGym benchmark, surpassing standalone offerings from major AI vendors.
The Numbers
Wiz's Project Atlas achieved 90.9% on CyberGym and has discovered more than 200 previously unknown vulnerabilities in open-source projects. Microsoft's MDASH scored 88.45% on the same benchmark and uncovered 16 new vulnerabilities in the Windows networking and authentication stack, including four critical remote code execution flaws.
How Multi-Model Works
Both systems abandon the premise that one frontier model can handle all security tasks. Instead, they route specific jobs to specialized agents optimized for different vulnerability classes.
MDASH deploys over 100 specialized agents, using MAI-Cyber-1-Flash for up to 90% of tasks and delegating complex cases to other models. Project Atlas assigns different security tasks to specialized AI agents across multiple model families, including Google's offerings.
"No single model is best at everything, and none stays state of the art for long," said Nir Ohfeld and Yuval Avrahami, Wiz's co-founders.
Cost and Performance
Microsoft claims its multi-model implementation reduces customer costs by 50% compared to relying exclusively on frontier models.
"As the models hand off between each other, they are not just able to deliver better performance than all of the other models combined, they do so at 50 percent of the cost," said Mustafa Suleyman, CEO of Microsoft AI.
Why It Matters
The industry is moving toward agentic security where AI doesn't just scan code but reasons through exploit chains and remediates issues. As frontier models become more expensive and specialized, vendors are shifting from relying on a single "best" model to orchestrating multiple models to balance performance, cost, and accuracy.
This shift demonstrates that the future of AI security lies in orchestration rather than just model scale. By routing tasks to specialized models, companies can achieve higher detection rates for zero-day vulnerabilities while significantly lowering the operational cost of continuous deep scanning.