Microsoft’s July 2026 Patch Tuesday is the largest in the company’s history: 570 vulnerabilities fixed in a single drop, up from 137 this time last year. That number is not an accident, and it is not a sign that Windows suddenly became dramatically more insecure. It is the direct output of MDASH — Microsoft’s internal AI vulnerability scanner. The same system that found most of those 570 flaws just entered public preview, and enterprise developers can now run it on their own code.
MDASH: The AI Vulnerability Scanner Behind the Numbers #
MDASH stands for Multi-Model Agentic Scanning Harness. The name matters: it is not a single model, it is an orchestration system for more than 100 specialized AI agents. Each agent has a distinct role in a four-stage pipeline.
In the Scan stage, auditor agents generate vulnerability hypotheses from the codebase. The Validate stage has debater agents argue for and against each finding — a deliberate adversarial check before anything reaches an engineer. Dedup removes overlapping signals through semantic analysis, and Prove dynamically constructs exploit inputs to confirm the bug is real before ever landing on a human’s desk.
The performance numbers are striking. In a controlled test with 21 deliberately planted vulnerabilities in a Windows device driver, MDASH found all 21 with zero false positives. On five years of confirmed cases in clfs.sys, it achieved 96% recall. On tcpip.sys, 100%. In May, Microsoft disclosed that MDASH had found 16 new vulnerabilities in the Windows networking and authentication stack — four of them critical remote-code-execution flaws. July’s record patch drop is what happens when that system runs continuously across the entire Windows codebase.
Microsoft is not locked to its own models here. MDASH uses a configurable panel that ranges from frontier models as heavy reasoners to more cost-effective models for high-volume passes. It includes third-party vulnerability discovery models and topped Anthropic’s Mythos on the CyberGym cybersecurity benchmark.
The Two Zero-Days That Need Immediate Action #
Of the 570 patches, two require the shortest triage window. Both were exploited before Microsoft released fixes.
CVE-2026-56155 is an Elevation of Privilege flaw in Active Directory Federation Services, carrying a CVSS score of 7.8. The root cause is insufficient access-control granularity. A low-privileged local attacker can escalate to administrator without any user interaction and with low attack complexity. Tenable’s CVE analysis notes this was found by Microsoft’s own DART team — which suggests active threat actor use, not just proof-of-concept.
CVE-2026-56164 hits SharePoint Server. CVSS 5.3, rated Moderate, but do not let the score anchor your triage. The attack vector is network-accessible, unauthenticated, and requires no user interaction. An attacker reaches SharePoint over the internet and elevates privileges. Affected versions include SharePoint 2019, Subscription Edition, and 2016. If your organization hosts SharePoint in any of those versions, this is not a “next maintenance window” patch.
Running MDASH on Your Own Code #
MDASH entered public preview in July 2026. The prerequisite is a Microsoft 365 E5 license or equivalent. From there, access works through Microsoft Security Exposure Management: find the preview toggle in the dashboard, install the Defender CLI extension, and follow the quick-start guide on Microsoft Learn.
For teams already on GitHub, the more compelling path is the GitHub connector. Configure it and MDASH scans trigger on pull requests, with findings surfacing as GitHub Advanced Security alerts inline on the PR itself. Azure DevOps users can gate builds and auto-generate work items. When a finding is confirmed, GitHub Copilot Autofix generates a remediation patch and assigns it in the workflow. One important constraint: MDASH output is a signal, not a verdict. The system flags high-confidence candidates — it does not bypass the human review step. Engineers still triage findings, validate impact, and review proposed fixes. Microsoft applies that same constraint internally: all AI-suggested code changes go through human review before release. The system finds the bugs faster. Humans still fix them.
What This Means for Patch Management #
The patch volume trajectory tells the real story. July 2025 produced 137 patches. June 2026 produced 200. July 2026 produced 570. Microsoft has said explicitly that patch volumes will keep increasing as AI finds more bugs faster than the monthly cycle was ever designed to handle.
This is not a Microsoft-specific problem. Anthropic’s Defending Code Harness, Strix’s AI pentesting platform, and now MDASH in public preview — every major vendor is moving toward proactive AI vulnerability scanning. The effect on defenders is cumulative: more findings discovered earlier means more patches arriving faster. Organizations treating monthly patch cycles as a comfortable cadence are already behind the curve. The real question is whether your patch management process can absorb 2x–4x volume spikes month over month.
The 570-patch record will likely not stand long. Microsoft’s own security team has said as much. The AI systems are running continuously. The discoveries keep coming.