Definition:
MDASH is Microsoft’s new AI-powered vulnerability-finding system that uses 100+ specialized AI agents working together (not just one model) to scan source code, debate whether a bug is real, and even prove it’s exploitable — all before a human ever looks at it.
Think of it as a team of AI security auditors, not just a single scanner:
MDASH Pipeline (5 stages):
┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐
│ PREPARE │ → │ SCAN │ → │ VALIDATE │ → │ DEDUP │ → │ PROVE │
│ Maps the │ │ 100+ │ │ "Debater" │ │ Collapse │ │ Build a │
│ attack │ │ specialized│ │ agents │ │ duplicate │ │ working │
│ surface │ │ auditors │ │ argue for │ │ findings │ │ exploit │
│ │ │ hunt bugs │ │ & against │ │ │ │ (PoC) │
└──────────┘ └──────────┘ └──────────┘ └──────────┘ └──────────┘Why it matters (results):
Benchmark Score CyberGym (1,507 real-world vulns) 88.45% — #1 on leaderboard 5 years of clfs.sys MSRC cases 96% recall 5 years of tcpip.sys MSRC cases 100% recall Private test driver (21 planted bugs) 21/21 found, zero false positives
Example:
One MDASH agent flagged a suspicious memcpy in ike_A.c that copied a security-realm identifier pointer. A second "debater" agent compared it against the correct pattern in ike_D.c (six files away), noticed the missing free, and concluded it was a real double-free bug. A third agent then constructed a proof — just two UDP packets triggering CVE-2026-33824, a critical RCE in the IKEv2 service running as LocalSystem. No single AI model could spot this alone; the orchestration of multiple agents made it possible.