Microsoft’s MDASH finds 16 Windows flaws with AI agents

What this is about

Microsoft introduced an internal security system codenamed MDASH on May 12, 2026. It is designed to review source code not with a single language model, but with more than 100 specialized AI agents that find suspicious paths, challenge each other and, where possible, prove the issue technically.

The concrete trigger is strong enough to be more than a product note: Microsoft says MDASH found 16 new vulnerabilities in the Windows networking and authentication stack. Four are rated critical, including remote-code-execution paths in components such as TCP/IP, IKEv2, Netlogon and DNSAPI. For everyday users, the meaning is simple: AI is not only being used to build attacks. It is now being used systematically to find the defects first.

What MDASH actually does

According to Microsoft, MDASH is a multi-stage review pipeline. First, the target codebase is prepared and indexed. Then auditor agents search for suspicious code paths. Other agents argue for and against each finding, check reachability and exploitability, deduplicate similar findings and, where the bug class allows it, try to prove the issue with a technical trigger.

The key difference from a normal chatbot is that the system does not rely on one answer. It uses different models and different roles. One agent searches, another challenges, a third tries to prove. Microsoft also describes domain plugins for specialist knowledge, such as kernel conventions, locking rules or component-specific data structures.

Why it matters

The numbers are unusually concrete. Microsoft cites 21 of 21 planted test vulnerabilities found with zero false positives in a private driver, 96 percent recall on historical MSRC cases in clfs.sys, 100 percent recall on historical cases in tcpip.sys and 88.45 percent on the public CyberGym benchmark of 1,507 real vulnerabilities.

For security teams, this matters because AI bug finding changes the patch economy. If large vendors find more vulnerabilities earlier, the short-term number of patches and CVEs rises. That is tiring for administrators, but better than discovering the same flaws first through attackers. It also shifts the competition: the winning factor is not one best model, but the full review pipeline of indexing, agent roles, proof and triage.

In plain language

Imagine a workshop inspecting a car before delivery. In the old model, one experienced mechanic looks under the hood and marks suspicious parts. MDASH is closer to a team of 100 inspectors: one looks for rust, one checks brake lines, one challenges false alarms, one drives a test lap, and only the repairs that truly matter end up on the list.

The point is not that every inspector is brilliant. The point is that the roles are separated cleanly and check each other.

A practical example

A company runs 12,000 Windows servers, including 1,400 with VPN or IPsec configurations. When a classic vulnerability bulletin arrives, the security team must ask: Which systems are affected, which services are actually active, and which patches deserve first priority?

For an MDASH-found issue like the described IKEv2 double-free, the practical question would be: Is there an IKEv2 responder policy, for example for RRAS VPN, DirectAccess, Always-On VPN or IPsec rules? If yes, the patch moves to top priority. If no, it remains important, but the immediate remote attack path is narrower. That kind of distinction makes security work faster and more honest.

Scope and limits

• Microsoft’s numbers come from its own tests and its own codebases. They are strong, but not automatically transferable to every programming language, company or legacy codebase.
• Finding more vulnerabilities does not automatically reduce risk. If patch processes are slow, a larger CVE wave can also overload teams.
• MDASH is not currently a generally available tool for every developer. Microsoft describes internal use and a limited private preview, not a broadly available open-source product.

SEO & GEO keywords

Microsoft MDASH, AI vulnerability discovery, Windows security, MSRC, Patch Tuesday, TCP/IP vulnerability, IKEv2 security, AI security agents, CyberGym benchmark, software supply chain, cybersecurity automation, vulnerability management