Adopt agentic AI vulnerability discovery (MDASH)
Implementation Effort: Medium – Evaluate and onboard an emerging preview capability and integrate its findings into existing triage workflows. User Impact: Low – Piloting MDASH and routing its findings is a security-team action in the Microsoft Defender portal; end users are not affected. Lifecycle Stage: Code
Preview capability. MDASH is an emerging Microsoft capability available in an expanded preview at the time of writing, and currently applies to GitHub-hosted code (via GitHub Code Security with native Microsoft Defender integration) — not Azure DevOps or GitLab. Treat this task as a forward-looking augmentation to evaluate and pilot — not yet a baseline control — and re-validate availability, supported platforms, naming, and licensing before relying on it in production.
Overview
Deterministic static analysis engines such as CodeQL match code against known vulnerability patterns and trace data flows, but they struggle with novel, multi-file, or logic flaws that no rule anticipates — and threat actors are increasingly using AI to discover exactly those flaws. To keep pace, evaluate and adopt agentic, AI-driven vulnerability discovery as a complementary layer over your existing code scanning.
Microsoft's multi-model agentic scanning harness (codename MDASH) was announced at Microsoft Build 2026 and is available in an expanded preview with native Microsoft Defender integration. Rather than relying on a single large language model or a fixed ruleset, MDASH orchestrates many specialized AI agents across multiple stages — code indexing, AI-driven scanning, adversarial "debate" to validate hypotheses, deduplication, and dynamic exploit proof generation. Because candidate findings are validated by attempting to trigger the flaw in a sandboxed environment before they are surfaced, the approach is designed to reduce false positives and prioritize bugs with demonstrated exploitability.
This is a distinct capability from traditional SAST, not a replacement for it. CodeQL and other deterministic scanners remain the baseline control: they are transparent, reproducible, run inline on every pull request, and integrate directly into the developer workflow. Agentic discovery complements that baseline by surfacing higher-order vulnerabilities that pattern-based tools miss, and by enriching findings with production risk context so security operations teams can triage them where they already work.
While the capability is in preview, a measured approach works best. Keep deterministic SAST as your baseline by implementing and operating CodeQL code scanning and custom query packs first, since agentic discovery augments, but does not substitute for, inline PR scanning. Pilot on high-value code, evaluating MDASH against your most security-sensitive services and internet-facing components rather than the entire estate, so you can assess signal quality before broad rollout. Route findings through existing triage: MDASH findings surface in the Microsoft Defender portal, so fold them into your established vulnerability-management and remediation workflows rather than creating a parallel process. Finally, track toward general availability by reassessing scope, licensing, and supported platforms as the capability moves from preview to GA, and promote it from a pilot to a standing control at that point.
Adopting agentic vulnerability discovery aligns with the Microsoft SDL practice to perform security testing and with NIST SSDF PW.7 and PW.8 (review and test executable and source code), strengthening the Assume breach posture by finding exploitable flaws before threat actors — increasingly AI-equipped themselves — can weaponize them.