CISO perspectives on managing vulnerability risks in the age of AI

What it is

Microsoft's MDASH is a "harness" layer wrapped around frontier AI models for vulnerability scanning — it controls how models access code, validates their outputs, and feeds findings into triage/remediation workflows. It's deployed internally across all Microsoft engineering groups and is now available to customers.

How it works

Raw AI model output for vuln-finding is non-deterministic — different runs or different models yield different results on the same codebase. The harness exists specifically to compensate: it mediates code access, checks/validates what the model claims to have found, and routes validated findings into existing triage and remediation pipelines rather than dumping raw model output on security teams. Microsoft also uses AI inside internal Red Team engagements to find control weaknesses faster.

Preconditions / Affected

Relevant to organizations running Microsoft on-premises software (expect continued high Patch Tuesday volumes — September 2026 hit near 1,000 disclosed vulnerabilities) and to any security team already overwhelmed by AI-generated finding volume. Applies whether or not an org has its own access to frontier AI models — the page explicitly says don't wait for that.

What to do

• Adopt a harness (e.g., MDASH) now to scan and remediate vulnerabilities in your own codebase — don't wait for frontier model access.
• Budget both token spend and human triage/remediation headcount for harness-driven output, since validated findings still require review capacity.
• Separately, deploy Microsoft Baseline Security Mode (BSM) for defense-in-depth; it's included in existing customer license agreements, supports scenario analysis to assess impact before rollout, and allows staged enablement with per-control exceptions.

Caveats

The page does not specify MDASH's architecture, which models it wraps, what validation checks it performs, or pricing/licensing terms for customer access — only that it exists and is now available. Non-determinism in AI scanning is stated as an ongoing limitation, not one the harness fully eliminates.