Microsoft's AI vulnerability scanner catches 10 critical flaws

Microsoft's experimental AI-powered vulnerability detection system has moved from research prototype to active security tool, discovering ten critical software flaws across Windows, Hyper-V, and Azure infrastructure that were patched in June 2026.

The system, codenamed MDASH, uses multiple specialized AI agents working in concert to scan proprietary code for security weaknesses. Rather than relying on a single model, it orchestrates different agents through a structured pipeline—each handling specific tasks like threat modeling, code analysis, and proof-of-concept generation.

According to a detailed technical post by Taesoo Kim, Vice President of Agentic Security at Microsoft, engineering teams across Windows, Azure, and identity systems now use MDASH as part of their standard security workflows. The system targets the deepest platform layers: kernel code, virtualization infrastructure, and Active Directory components that require understanding complex trust boundaries and object lifetime rules.

The June 2026 discoveries

The ten vulnerabilities MDASH identified span multiple attack classes. Three involve remote code execution flaws in Hyper-V, with CVSS scores ranging from 8.2 to 8.4. The system also caught a severity 9.8 use-after-free bug in the Windows kernel and a 9.8 integer overflow in HTTP.sys—both rated critical for remote code execution risk.

Other findings include elevation-of-privilege vulnerabilities in the Windows DNS client, information disclosure issues in DHCP client components, and remote code execution flaws in Active Directory Domain Services and Remote Desktop Client. Each was identified before any known exploitation occurred.

Microsoft first reported these details in its security blog.

Integration with existing tools

MDASH plugs directly into Microsoft's development infrastructure rather than operating as a standalone scanner. Validated findings appear as code scanning alerts in GitHub Advanced Security, surface inline on pull requests, and flow into Azure DevOps pipelines where they can gate builds. The same findings feed into Microsoft Defender, where security teams prioritize them alongside threat intelligence and runtime signals.

This integration means discovered vulnerabilities enter the standard remediation workflow—assigned to an owner, tracked through a pull request, and fixed as part of normal engineering cycles.

Benchmark performance and limitations

The latest version of MDASH achieved 96.5% accuracy on CyberGym, an industry benchmark built from 1,507 real-world vulnerabilities. The system missed 52 cases, with 65% of failures occurring in the proof-of-concept generation stage. These failures typically involved targets requiring highly structured binary inputs or complex build environments.

Microsoft attributes recent improvements to better scoping logic, more comprehensive threat modeling, and smarter routing between specialized agents. The company tested newer foundation models including GPT-5.5 on previously failed cases, solving an additional 36.5% and projecting potential accuracy of 97.8%.

Why it matters

Traditional security reviews happen at fixed points in the development cycle, creating windows where vulnerabilities exist in shipped code before discovery. AI-driven continuous scanning compresses that window, particularly for complex platform code that demands significant manual effort to audit. By catching critical flaws in Windows kernel and hypervisor code before attackers find them, Microsoft demonstrates how AI agents can extend—not replace—human security expertise at enterprise scale. The system's integration with existing DevSecOps tools suggests a model for making AI security capabilities practical rather than experimental.

Microsoft detailed these findings and technical approach in a June 17, 2026 post on the company's security blog.