Microsoft has introduced a groundbreaking multi-model artificial intelligence (AI) system known as MDASH, aimed at enhancing vulnerability discovery and remediation processes on a large scale. Currently, this innovative solution is undergoing testing with select customers as part of a limited private preview.
MDASH: A New Era in Vulnerability Management
MDASH, which stands for multi-model agentic scanning harness, is crafted to function as a model-agnostic system. It employs tailored AI agents for various classes of vulnerabilities, allowing for the autonomous discovery, validation, and demonstration of exploitable defects within intricate codebases, such as those found in Windows.
“Unlike traditional single-model approaches, MDASH orchestrates over 100 specialized AI agents across a diverse ensemble of both frontier and distilled models. This enables the system to discover, debate, and conclusively prove exploitable bugs from start to finish,” explained Taesoo Kim, Microsoft’s vice president of agentic security.
The architecture of MDASH is envisioned as a “structured pipeline” that processes a codebase to yield validated and proven findings through a series of methodical actions. The process begins with an analysis of the source code to construct a threat model and assess the attack surface. It then employs specialized “auditor” agents to scrutinize candidate code paths for potential issues, followed by “debater” agents that validate these findings. The system groups semantically equivalent findings and ultimately proves the existence of the identified vulnerabilities.
MDASH operates on a customizable panel of models, utilizing state-of-the-art (SOTA) models for reasoning, distilled models for high-volume validation, and an additional SOTA model for independent counterpoint analysis. Microsoft elaborated, “Disagreement between models serves as a significant indicator: when an auditor flags a potential issue that the debater cannot refute, the credibility of that finding increases.” Each stage of the pipeline is designed with distinct roles, prompt regimes, tools, and criteria for halting processes.
Redmond has noted that the specialized agents are built upon historical common vulnerabilities and exposures (CVEs) and their corresponding patches. The architecture is also designed to ensure portability across different model generations.
In its initial tests, MDASH successfully identified 16 vulnerabilities that were addressed in the recent Patch Tuesday release. These vulnerabilities affect the Windows networking and authentication stack and include two critical flaws that could enable remote code execution:
- CVE-2026-33824 (CVSS score: 9.8) – A double-free vulnerability in “ikeext.dll” that could permit an unauthenticated attacker to send specially crafted packets to a Windows machine with Internet Key Exchange (IKE) version 2 enabled, resulting in remote code execution.
- CVE-2026-33827 (CVSS score: 8.1) – A race condition vulnerability in Windows TCP/IP (“tcpip.sys”) that allows an unauthorized attacker to send a specially crafted IPv6 packet to a Windows node where IPSec is enabled, leading to potential remote code execution exploitation.
The announcement of MDASH coincides with the launch of Anthropic’s Project Glasswing and OpenAI’s Daybreak, both of which are AI-driven cybersecurity initiatives designed to expedite the discovery, validation, and remediation of vulnerabilities before they can be exploited by malicious actors.
Kim emphasized the strategic significance of this development, stating, “AI vulnerability discovery has transitioned from a research curiosity to a production-grade defense mechanism at the enterprise level. The sustainable advantage lies in the agentic system surrounding the model rather than in any single model itself.”
