Microsoft has shown how artificial intelligence can be used to hunt down Windows flaws before attackers get to them. Its MDASH system uncovered 16 vulnerabilities, including four critical bugs that could have opened direct access into enterprise networks.
What makes the discovery significant is where several of the flaws were found. Many sat inside core Windows components and could be reached over the network without credentials, which would have made them especially dangerous if left unpatched.
Microsoft says all 16 issues were fixed in Patch Tuesday on May 12. Satya Nadella then shared the news on X the following day.
How MDASH works
MDASH stands for Multi-model Agentic Scanning Harness. It was built by Microsoft’s Autonomous Code Security team, with several members who previously worked on Team Atlanta, the DARPA AI Cyber Challenge winner that took home $29.5 million.
The system does not rely on a single model to scan code. Instead, it runs more than 100 specialized agents across a mix of frontier and distilled models, with each agent handling a different task such as finding defects, testing findings, and proving whether a bug is actually exploitable.
Microsoft only shows the results to human engineers after the full verification process is complete. That approach matters because some Windows weaknesses require reasoning across multiple code paths, something standard scanners often miss.
Where the vulnerabilities were found
Microsoft said the 16 vulnerabilities were spread across the Windows TCP/IP stack, the IKEEXT IPsec service, HTTP.sys, Netlogon, the Windows DNS Client, and the Telnet client. Ten of them were in kernel mode, and most could be reached over the network without authentication.
Two of the four critical issues drew the most attention. CVE-2026-33827 was located in tcpip.sys and was triggered by a specially crafted IPv4 packet, while CVE-2026-33824 was a pre-authentication double-free in the IKEEXT service that could be reached through UDP port 500 on machines running RRAS VPN, DirectAccess, or Always-On VPN.
Both of those bugs could lead to LocalSystem execution. The other two critical vulnerabilities were found in Netlogon and the Windows DNS Client, and each carried a CVSS score of 9.8.
Why ordinary scanners can miss them
Microsoft says these were not the kinds of bugs that usually appear in standard scanner results. The tcpip.sys flaw required reasoning across three concurrent code paths that all freed the same object, while the IKEEXT case involved six different source files.
That complexity helps explain why a single-model approach is often not enough when the goal is to find bugs that are truly exploitable. In practice, the security value came not just from detection, but from being able to verify the exploit path before the finding was handed to humans.
Performance and current availability
MDASH also performed strongly in UC Berkeley’s CyberGym benchmark, which contains 1,507 real-world vulnerability reproduction tasks. It scored 88.45 percent, placing it at the top of the public leaderboard ahead of Anthropic Mythos Preview at 83.1 percent and OpenAI’s GPT-5.5 at 81.8 percent.
In private testing against an unreleased Windows driver codebase called StorageDrive, MDASH found all 21 deliberately planted vulnerabilities with no false positives. On confirmed MSRC cases spanning five years in clfs.sys and tcpip.sys, its recall reached 96 percent and 100 percent.
For now, MDASH remains in a private preview limited to a small group of enterprise customers. Microsoft says wider availability is expected in the coming months, and the system is model-agnostic, which means the underlying model can be swapped without rebuilding the pipeline.
The move comes as other major AI security efforts also advance under limited access, including Anthropic’s Project Glasswing and OpenAI’s Daybreak. At the same time, Google has confirmed the first known zero-day developed with AI for a mass exploitation campaign targeting 2FA bypass on widely used web admin tools.
