On May 12, Microsoft's Autonomous Code Security team published a result that should make every developer building single-model pipelines uncomfortable: their system MDASH scored 88.45% on the CyberGym vulnerability benchmark, outpacing Anthropic's Mythos Preview (83.1%) by five points and OpenAI's GPT-5.5 (81.8%) by nearly seven. The kicker? MDASH doesn't use one model—it coordinates more than 100 specialized agents across a pipeline of frontier and distilled models working in concert.

What Makes MDASH Different

MDASH stands for Multi-model Agentic Scanning Harness, and the architecture is deliberately unapologetic about its philosophy: composition beats scale when tasks decompose into distinct phases requiring different reasoning patterns. Frontier models handle heavy lifting as reasoners. Distilled models process high-volume filtering. A separate state-of-the-art model acts as an independent counterpoint. Taesoo Kim, Microsoft's VP of Agentic Security, put it plainly: 'The harness does the work, and the model is one input.' That's the thesis in a single sentence—in 72 hours, your best model can vanish or change pricing; your orchestration layer persists.

The Five-Stage Pipeline

MDASH runs through Prepare, Scan, Validate, Dedupe, and Prove stages. Auditors flag suspicious code paths with hypotheses but no validation. Debaters then argue each finding's reachability and exploitability—disagreement between cohorts becomes signal rather than noise. A deduplication layer collapses semantically equivalent findings via patch-based grouping before expensive prover agents construct working exploits using AddressSanitizer for C/C++ targets. This specialization catches cross-file ownership bugs that collapse into silence when a single model processes each function in isolation—the exact class of vulnerabilities that cause real-world breaches.

The CVE That Proves the Point

The benchmark numbers are self-reported (GeekWire's Todd Bishop flagged this), so treat them as directional rather than gospel. But MDASH didn't just win a contest—it found 16 previously unknown Windows vulnerabilities on its way to Patch Tuesday, including CVE-2026-33824: a double-free in the IKEEXT service reachable remotely over UDP port 500 by an unauthenticated attacker. Four of those findings were critical remote code execution flaws across kernel-mode and user-mode components. That's not a demo—that's production output. By Build 2026 on June 2, MDASH had climbed to 96.55%, gaining roughly ten percentage points in under three weeks through model-panel refinements rather than architectural rewrites.

Why This Matters Beyond Security

The CyberGym benchmark—developed by UC Berkeley researchers across 1,507 tasks from 188 open-source projects—tests exactly the kind of work that exposes single-model ceilings: cross-file reasoning, multi-step validation, and proof construction. No single model excels at all three simultaneously. MDASH's design principle transfers directly to other domains: when your workflow has distinct phases requiring different cognitive approaches, build a pipeline of specialists optimized for each stage rather than prompting one generalist harder.

Key Takeaways

  • Composition beats scale on complex multi-step tasks—no single model can excel at discovery, validation, and proof simultaneously
  • Model-agnostic architecture means swapping in better models requires changing config files, not rewriting pipelines
  • Ensemble disagreement is signal: build systems that surface conflicting agent views rather than suppress them
  • The real-world CVE output validates the approach more than benchmark percentages ever could

The Bottom Line

MDASH answers a question the industry has been dancing around for two years: what can you build when you stop searching for one model to rule them all? The answer, apparently, is the top of the leaderboard—and sixteen critical Windows vulnerabilities that would have shipped unpatched otherwise. If you're still building single-model pipelines for complex workflows, you're not optimizing—you're leaving performance on the table.