Microsoft MDASH AI found 16 Windows bugs fixed in Patch Tuesday

Microsoft tested MDASH, a multi-model agentic scanning harness, in a limited private preview with select customers and reported that it uncovered 16 vulnerabilities fixed in this month’s Patch Tuesday. The system is designed to find, validate and demonstrate exploitable defects in large codebases such as Windows.

MDASH ingests a codebase, analyzes it to build a threat model and map an attack surface, then runs targeted agents over candidate code paths. Auditor agents flag potential issues, a separate set of debater agents attempts to validate or refute those flags, and prover agents try to demonstrate exploitability. Semantically similar findings are grouped before being passed to provers for evidence building.

Each stage uses its own prompt regimes, tools and stop criteria to reduce false positives and produce evidence for actionable findings, Microsoft explained. The platform is model-agnostic and orchestrates an ensemble of state-of-the-art models and smaller distilled models. Larger models handle complex reasoning, distilled models run high-volume validation passes, and a separate high-capacity model provides an independent counterpoint. Microsoft noted that disagreement between agents can be informative: if an auditor flags an issue and a debater cannot refute it, the finding’s credibility increases.

In tests, MDASH surfaced 16 of the vulnerabilities that Microsoft fixed in the recent Patch Tuesday release. The issues spanned the Windows networking and authentication stack and included two high-severity remote code execution flaws. CVE-2026-33824 (CVSS 9.8) is a double-free vulnerability in ikeext.dll that could allow an unauthenticated attacker to send specially crafted packets to a device running Internet Key Exchange (IKE) version 2 and achieve remote code execution. CVE-2026-33827 (CVSS 8.1) is a race condition in the Windows TCP/IP driver (tcpip.sys) that may be triggered by a specially crafted IPv6 packet on nodes where IPSec is enabled, also enabling remote code execution.

Microsoft said the agents were built from patterns seen in past vulnerabilities and patches, and that the architecture is portable across model generations so future models can be swapped into the pipeline without changing its structure. Taesoo Kim, vice president of agentic security at Microsoft, wrote: “Unlike single-model approaches, the harness orchestrates more than 100 specialized AI agents across an ensemble of frontier and distilled models to discover, debate, and prove exploitable bugs end-to-end.”

MDASH remains in a limited private preview while Microsoft continues testing and refining the system with partner customers. The company has not disclosed a public release date and said the tool is intended to scale automated vulnerability discovery and remediation across large, complex codebases.