MDASH — Microsoft Multi-Model Agentic Scanning Harness

Sources: Microsoft Security Blog — MDASH announcement (May 12, 2026) · Private Preview Signup (aka.ms)

MDASH (codename — multi-model agentic scanning harness) is Microsoft’s agentic vulnerability discovery and remediation system, built by Microsoft’s Autonomous Code Security (ACS) team in collaboration with Microsoft Windows Attack Research and Protection (WARP). It orchestrates 100+ specialized AI agents across an ensemble of frontier and distilled models — auditors, debaters, dedup agents, and provers — to find, validate, and prove exploitable vulnerabilities end-to-end. First publicly announced May 12, 2026; in limited private preview as of mid-May 2026; product naming uses “codename MDASH” — the GA product name is not yet disclosed.

Pipeline Architecture

Five-stage pipeline; targeting / validation / dedup / prove stages are model-agnostic by construction:

StageRoleOutput
PrepareIngests source target; builds language-aware indices; draws attack-surface and threat models from past commits.Indices, attack-surface model
ScanSpecialized auditor agents reason over candidate code paths.Candidate findings with hypotheses and evidence
ValidateDebater agents argue for and against each finding’s reachability and exploitability.Surviving findings with strengthened rationale
DedupCollapses semantically equivalent findings (patch-based grouping).Distinct vulnerability candidates
ProveConstructs and executes triggering inputs where the bug class admits (e.g., ASan in C/C++).Validated PoC artifacts

Three Architectural Properties

  1. Ensemble of diverse models. SOTA as heavy reasoner; distilled models as cost-effective debater for high-volume passes; a second separate SOTA model as independent counterpoint. Disagreement between models is itself a signal. Microsoft does not disclose which specific models occupy these roles.

    Cross-model disagreement as a validation signal is the field’s convergent default rather than a Microsoft architectural bet. Semgrep’s July 2026 survey records a second independent agent attempting to falsify each finding in four of nine open-source harnesses — Cloudflare’s Phase 3, VVAH’s stage S6, Trail of Bits’ fp-check, and defending-code-harness’s fresh-container grader — and states the accuracy gain is largest where a different model attempts the disproof.1 What stays specific to MDASH is scale and specialization: 100+ role-bound agents, distilled models occupying the high-volume debater slot, and domain plugins such as the CLFS prover that inject context the foundation models cannot see.

  2. Specialized agents. 100+ agents, each with its own role, prompt regime, tools, and stop criteria. Auditors do not reason like debaters; debaters do not reason like provers. Constructed through deep research with past CVEs and their patches.

  3. End-to-end pipeline with extensible plugins. Domain plugins inject context the foundation models cannot see — kernel calling conventions, IRP rules, lock invariants, IPC trust boundaries, codec state machines, custom CodeQL databases. The CLFS proving plugin is a worked example, embedding on-disk container layout + block-validation sequence + in-memory state machine to construct triggering log files for candidate findings.

Public Benchmark Results

BenchmarkResultProvenance
StorageDrive (21 planted vulns, private Microsoft interview driver)21/21 found, 0 false positivesMicrosoft-disclosed
clfs.sys historical recall (5 years, 28 MSRC cases)96%Microsoft-internal
tcpip.sys historical recall (5 years, 7 MSRC cases)100%Microsoft-internal
CyberGym public leaderboard (1,507 real-world vuln-repro tasks, level 1)88.45%Top score, ~5 points above #2 (83.1%)
May 2026 Patch Tuesday16 new CVEs10 kernel-mode, 6 usermode; 4 Critical RCEs

Of the five results above, only the CyberGym number is independently verifiable; the others are Microsoft-internal but anchored to a defensible ground truth (MSRC case database).

Positioning

MDASH carries two wiki scope axes:

  • ai-in-sec-defense (primary): Microsoft’s defender-AI capability at the AppSec / vulnerability-research layer, distinct from Security Copilot at the SOC layer. Second sourced anchor for Frontier AI for Vulnerability Discovery after XBOW/Mythos; the convergence with XBOW — two vendors, opposite sides of the stack, arriving at “the harness does the work, the model is one input” — is itself the load-bearing observation.
  • sec-of-ai (tertiary): MDASH is itself an agentic system; the CMM questions about agent identity, action authority, and audit apply to MDASH’s 100+ agents.

Convergence with XBOW + Mythos

XBOW + MythosMDASH
OrientationOffensive (live-site pentest)Defensive (developer-side audit)
Model strategySingle best frontier model, strong harnessEnsemble (SOTA + distilled + second-SOTA counterpoint)
ValidationLive-site interaction harnessDebate + dedup + automated prove (PoC construction)
TargetsLive web applicationsNative Windows components (tcpip.sys, ikeext.dll, etc.)
Key insight (paraphrased)“A model is a brain without a body""The harness does the work; the model is one input”
Public benchmarkXBOW’s internal web exploit benchmarkCyberGym public leaderboard
PricingMythos ~5× Opus at GANot disclosed; private preview

The two systems are architecturally distinct but converge on the same architectural argument — orchestration outperforms model choice, and validation infrastructure determines real-world utility.

Distribution

  • Status: limited private preview as of May 2026.
  • Users: Microsoft security engineering teams (production); a small set of customers (preview).
  • Signup: aka.ms/AI-drivenScanningHarness.
  • GA timeline / pricing / SKU positioning: not yet disclosed.

Open Questions

  • Which models? Microsoft attributes the ensemble only as “generally available AI models,” with no further detail disclosed publicly. Anthropic’s Mythos is a plausible SOTA-reasoner candidate; OpenAI GPT and Microsoft-internal models are alternatives. Microsoft’s silence is conspicuous.
  • GA productization: standalone product, feature within Defender / Security Copilot, or service offering via consulting? The post does not commit.
  • Customer co-pilot pattern: the post says customers “test” MDASH — implies hosted-service model rather than on-prem deployment. To be confirmed.
  • CyberGym configuration: 88.45% is at level 1 (vulnerable source + high-level description supplied). Performance on higher-difficulty levels (blind discovery) would be a stronger signal.
  • Relationship to MITRE ATLAS or other vulnerability taxonomies: not addressed.

See Also

Notes

Footnotes

  1. Semgrep — Comparing open source AI code security harnesses, July 2026 (no day-level date exposed; author not named). The four adversarial-validation instance names and the different-model observation are human-written. Summarized at OSS AI Security Harness Comparison.