Microsoft Security · 2026-05-12 · major
Microsoft MDASH — Multi-Model Agentic Scanning Harness Finds 16 Windows Flaws Including Four Critical RCEs, Tops CyberGym at 88.45%
Microsoft's Autonomous Code Security team unveils MDASH, an ensemble of 100+ agents across frontier and distilled models that scans, debates, and proves bugs. Surfaced 16 new CVEs in Windows fixed in May Patch Tuesday.
Microsoft's new agentic vulnerability discovery system orchestrates 100+ AI agents across frontier and distilled models, and just topped CyberGym.
Key specs
| Cyber gym score | 88.45% |
|---|---|
| Msrc recall (clfs.sys, 5y) | 96% |
| Msrc recall (tcpip.sys, 5y) | 100% |
| Planted bug test | 21/21, 0 FPs |
| New cves found | 16 |
| Critical rces | 4 |
| Agents in ensemble | 100+ |
What is it?
MDASH stands for multi-model agentic scanning harness. Built by Microsoft's Autonomous Code Security team, it is a model-agnostic system that uses bespoke agents for different vulnerability classes to autonomously discover, validate, and prove exploitable defects in complex codebases like Windows. On its first production run it surfaced 16 previously unknown Windows vulnerabilities, including four critical remote code execution flaws in the Windows kernel TCP/IP stack and the IKEv2 service — all shipped in this month's Patch Tuesday.
How does it work?
The harness pipelines five sequential stages — Prepare, Scan, Validate, Dedup, and Prove — across more than 100 specialized agents drawn from an ensemble of frontier and distilled models. Auditor agents flag candidate bugs in source code; debater agents argue findings for or against; a dedup stage collapses semantically equivalent results; and a Prove stage constructs and executes triggering inputs before anything reaches a human engineer. Microsoft says the ensemble approach beats single-model agentic scanners because different models catch different bug classes.
Why does it matter?
MDASH lands in the same lane as Anthropic's Project Glasswing and OpenAI's Daybreak, but Microsoft is publishing the first end-to-end benchmark numbers any of them has shared. The 88.45% CyberGym score is roughly five points ahead of the next-best published system; 96–100% recall on five years of historical MSRC cases means it can rediscover bugs that took human teams years to chase. Microsoft has telegraphed that AI-assisted discovery will keep inflating Patch Tuesday counts — May 2026's wave is already 138 CVEs, the largest of the year.
Who is it for?
security engineers, vulnerability researchers, blue teams
Try it
Request limited private preview at aka.ms/AI-drivenScanningHarness