Microsoft’s latest AI-driven security tool has identified 16 previously unknown vulnerabilities in Windows, including four critical remote code execution flaws, ahead of any hacker exploitation. This new system, known as MDASH, operates by deploying over 100 specialized AI agents to scan codebases intensively, which could mark a turning point for software security.

How MDASH Detects Vulnerabilities

Unlike traditional scanners, MDASH functions like a full security team packed into software. It processes scans through multiple stages such as preparation, scanning, validation, deduplication, and proofing, employing distinct AI models tailored for each phase rather than relying on a single AI model. This approach delivered an 88.45% score on CyberGym’s public benchmark, outpacing competitors by about five percentage points. Internal tests showed even higher precision: a 96% detection rate for past clfs.sys vulnerabilities and perfect accuracy for tcpip.sys, with no false positives on planted bugs.

Microsoft also unveiled a companion model, MAI-Cyber-1-Flash, which scored 96% on the same benchmark on its own. The 16 vulnerabilities MDASH uncovered were patched in May 2026’s Patch Tuesday update, highlighting how these discoveries have real-world impact across billions of devices.

Implications for Crypto and Cybersecurity

The architecture of MDASH supports major programming languages and allows integration of domain-specific plugins, making it adaptable for auditing complex environments like Solidity or Rust smart contracts. This adaptability suggests a potential role in securing crypto and Web3 projects where code vulnerabilities can lead to significant financial losses.

Currently, MDASH is available in limited preview through Azure's commercial cloud, with partners such as CyberOne already engaged. Microsoft's achievement sets fresh performance standards within cybersecurity, emphasizing tangible outcomes over research alone. This development ramps up competition for dedicated security firms, challenging them to meet or exceed the new AI-powered benchmark.