Microsoft MDASH Shows How Multi-Agent Security Systems Can Improve Vulnerability Discovery

Microsoft's Multi-Model Agentic Scanning Harness, or MDASH, is an AI security system designed to coordinate multiple models and agents for vulnerability discovery. The system has recorded an 88.45% score on the public CyberGym leaderboard, according to Microsoft's reported results, while separate evaluations identified Windows vulnerabilities and measured recall against historical Microsoft Security Response Center cases.

The important development is the agentic security pipeline itself, not an independently documented MAI-Cyber-1-Flash model. Microsoft's official MDASH security announcement describes a multi-model approach to automated security research. That framing matters for enterprises evaluating AI-assisted defense: the performance of an AI security workflow can depend on how models, tools, verification steps, and agents are orchestrated, rather than on one model alone.

Microsoft reported that MDASH found 16 CVEs during a Patch Tuesday evaluation. In a test driver containing 21 planted vulnerabilities, it found all 21 with zero false positives. The company also reported 96% recall against five years of MSRC cases for clfs.sys, plus 100% recall for tcpip.sys.

What Microsoft's MDASH results measure

MDASH is a multi-model agentic scanning harness, meaning it uses a coordinated pipeline rather than treating vulnerability research as a single prompt-and-response task. That distinction is significant because software security testing often requires several linked activities: analyzing code or binaries, generating hypotheses, attempting exploitation paths, checking findings, and filtering results that do not hold up under validation.

CyberGym is one measure of how well an automated security system performs in a benchmark environment. Microsoft's reported public leaderboard score is 88.45%. The 96% figure in the supplied research refers instead to recall against historical MSRC cases for clfs.sys; it is not a corroborated CyberGym result.

These measurements should not be collapsed into one broad claim about all security tasks. A leaderboard score, planted-vulnerability test, and recall measured against historical cases answer different questions. Together, however, they indicate that Microsoft is evaluating MDASH across benchmark, controlled-test, and historical-case contexts.

Why multi-agent security workflows matter to enterprises

For security teams, a useful AI system is not simply one that identifies possible flaws. It must also fit into a defensible workflow that helps researchers prioritize, investigate, reproduce, and validate findings. The reported zero-false-positive result on the planted-vulnerability driver is particularly relevant because false positives can consume scarce security engineering time.

A multi-model architecture can also make governance more important, not less. Enterprises adopting agentic security systems need clarity on which models and tools are invoked, what data they can access, how outputs are logged, and where human review is required. The operational questions include:

• Access controls: which repositories, binaries, environments, and security tools agents can reach.
• Validation gates: how candidate vulnerabilities are reproduced and checked before escalation.
• Auditability: whether teams can trace a reported finding to the evidence and actions that produced it.
• Human accountability: who approves remediation, disclosure, and changes to production systems.

The available research does not provide public MDASH pricing or commercial availability details. Organizations should therefore separate the demonstrated technical results from procurement assumptions. A benchmark result can establish that a security pipeline is technically notable, but it does not by itself define how external customers can deploy it or under what commercial terms.

Organizations considering agentic vulnerability research can work with Scalevise on AI architecture, workflow automation, and governance design that connects automated analysis with controlled human validation and existing security processes.

Frequently Asked Questions

What is Microsoft MDASH?

MDASH is Microsoft's Multi-Model Agentic Scanning Harness, a security system that coordinates multiple models and agents for vulnerability discovery and related security research tasks.

What was MDASH's CyberGym score?

Microsoft reported that MDASH achieved an 88.45% score on the public CyberGym leaderboard.

Did MDASH achieve 96% on CyberGym?

No. The supplied research identifies 88.45% as the reported public CyberGym score. The 96% figure refers to recall against five years of MSRC cases for clfs.sys.

What vulnerability findings did Microsoft report for MDASH?

Microsoft reported 16 CVEs identified in a Patch Tuesday evaluation. It also reported finding 21 planted vulnerabilities in a test driver with zero false positives.

Are MDASH pricing and availability public?

The supplied research does not provide public pricing or availability details for MDASH.

Conclusion

MDASH's reported results make a clear case for evaluating AI security as an orchestrated system of models, agents, tools, and validation steps. Its 88.45% CyberGym score, reported CVE findings, and component-specific recall figures provide meaningful evidence of that approach, while enterprises still need practical clarity on deployment, controls, and commercial access.