Microsoft’s MDASH AI Finds 16 Windows Security Flaws Before Hackers Could Exploit Them
Microsoft’s new AI-powered security system has already delivered a major win for Windows users and enterprise networks. The system, known internally as MDASH, discovered 16 Windows vulnerabilities before attackers were able to exploit them. Even more significantly, four of those flaws were rated critical and could have allowed remote code execution in certain environments.
All 16 vulnerabilities were fixed as part of Microsoft’s May 12 Patch Tuesday security update, giving administrators an urgent reason to apply the latest Windows patches as quickly as possible.
MDASH, short for Multi-model Agentic Scanning Harness, was developed by Microsoft’s Autonomous Code Security team. The project brings together advanced AI models and specialized security agents designed to search through complex code, verify potential weaknesses, and determine whether a bug can actually be exploited.
Unlike traditional vulnerability scanners, MDASH does not simply flag suspicious code and pass it to a human reviewer. Instead, it uses more than 100 specialized AI agents working across multiple models. Some agents search for possible vulnerabilities, others challenge those findings to reduce false positives, and another stage attempts to generate proof that the issue is exploitable. Only after that does a Microsoft security engineer review the result.
That layered approach appears to be paying off.
The vulnerabilities discovered by MDASH affected several major Windows components, including the Windows TCP/IP stack, IKEEXT IPsec service, HTTP.sys, Netlogon, Windows DNS, and the Telnet client. Ten of the flaws were found in kernel-mode code, which is especially sensitive because successful exploitation can give attackers deep control over a system.
Most of the bugs could be reached over a network without requiring login credentials, increasing their risk for businesses running exposed or poorly segmented systems.
Two critical vulnerabilities were especially serious.
One, tracked as CVE-2026-33827, was found in tcpip.sys and could be triggered using specially crafted IPv4 packets. Another, CVE-2026-33824, was a pre-authentication double-free vulnerability in the IKEEXT service. That flaw was reachable through UDP port 500 on systems using RRAS VPN, DirectAccess, or Always-On VPN.
Both vulnerabilities could allow attackers to execute code with LocalSystem privileges, one of the highest levels of access on Windows.
Two additional critical flaws were found in Netlogon and the Windows DNS Client, each receiving a CVSS score of 9.8. A score that high indicates a severe vulnerability that administrators should prioritize immediately.
Microsoft says these were not simple bugs that conventional tools would normally catch. The tcpip.sys issue required the AI system to reason across three separate concurrent code paths that were all freeing the same object. The IKEEXT vulnerability was spread across six different source files, making it difficult to detect through basic automated scanning or single-model code review.
This is where MDASH appears to stand apart. Complex vulnerabilities often hide in the relationships between different parts of a codebase. A tool looking at only one function, one file, or one execution path can easily miss the bigger picture. MDASH was designed to connect those dots.
Microsoft also tested MDASH against CyberGym, a UC Berkeley benchmark made up of 1,507 real-world vulnerability reproduction tasks. The system scored 88.45%, placing it at the top of the public leaderboard. For comparison, Anthropic’s Mythos Preview model scored 83.1%, while OpenAI’s GPT-5.5 scored 81.8%.
In another private test, Microsoft used a Windows driver codebase called StorageDrive that had not been publicly released. MDASH found all 21 planted vulnerabilities with zero false positives. Microsoft also tested it against five years of confirmed security cases in clfs.sys and tcpip.sys, where it achieved 96% recall and 100% recall respectively.
One of MDASH’s biggest advantages is that it is model-agnostic. Microsoft can replace or upgrade the AI models behind the system without rebuilding the entire pipeline. As newer AI models become available, MDASH can potentially become faster, more accurate, and more capable at detecting subtle security flaws.
For now, MDASH is available only in a limited private preview with a small number of enterprise customers. Microsoft is expected to expand access in the coming months.
The timing is important. AI is rapidly changing both sides of cybersecurity. Defenders are using artificial intelligence to find vulnerabilities before attackers do, while attackers are also exploring AI-assisted exploit development. That makes tools like MDASH increasingly important for protecting operating systems, cloud infrastructure, and enterprise networks.
For Windows administrators, the takeaway is clear: apply the May 12 Patch Tuesday updates as soon as possible. Several of the patched flaws were remotely reachable, required no authentication, and affected core Windows networking components. Delaying updates could leave systems exposed to vulnerabilities that are now publicly known.
MDASH shows how AI could reshape software security by identifying dangerous flaws earlier in the development and patching process. If Microsoft can scale this approach across more Windows components and enterprise products, AI-driven vulnerability discovery may become a key defense against the next wave of cyberattacks.
