Microsoft this month quietly opened its MDASH AI-powered vulnerability scanner to select customers, while separately gearing up to launch a commercial version of the technology as early as July. The moves mark a rapid progression from internal research—the system was disclosed just two months ago—to hands-on customer testing, as the company pushes AI to the forefront of its security business.
The Private Preview Is Live, and the Commercial Product Has a Name
Microsoft’s Multi-Model Agentic Scanning Harness—MDASH for short—entered private preview inside Microsoft Security Exposure Management in July 2026. This is the same system that, in a May blog post, the company credited with uncovering 16 previously unknown vulnerabilities in Windows networking and authentication components. Four of those were critical remote-code-execution flaws, including issues in the Windows kernel TCP/IP stack and IKEv2. Fixes shipped as part of the May Patch Tuesday release, giving the research real-world teeth.
Now, approved testers can run MDASH scans through a Defender CLI or a GitHub connector, review findings in the Microsoft Defender portal, and use those findings to prioritize code-security risk. Microsoft describes a workflow where human security engineers remain firmly in control, with AI findings feeding into existing triage and patch pipelines—not an autonomous patch-bot.
In parallel, The Information reported on July 16 that the company is preparing a standalone product codenamed Project Perception. The service would use a model router to coordinate tasks among Microsoft, OpenAI, and Anthropic models, reserving expensive reasoning for deep exploitability analysis while cheaper models handle bulk scanning and deduplication. The target launch window, according to that report: as soon as July.
What Changed for You: It Depends on Your Role
For Everyday Windows Users
You won’t click a button to run MDASH on your laptop, but you will feel its effects. Every vulnerability the system surfaces for Microsoft’s own products means a patch for your machine. The 16 flaws already caught—four of them critical—illustrate the potential for faster, more proactive fixes. But more findings also mean a busier Patch Tuesday. Keeping automatic updates turned on remains your best defense. If you’re on a managed device, expect administrators to push updates more frequently as AI narrows the gap between discovery and disclosure.
For IT Administrators and Security Teams
The private preview is the immediate opportunity. If your organization is part of the program, you can now integrate MDASH with your existing Defender and GitHub workflows. That’s a practical step, not a theoretical one: connect the GitHub connector, point it at a repository, and start receiving AI-triaged findings in the portal you already use.
But treat those findings as sophisticated triage, not a final verdict. Microsoft hasn’t released detailed false-positive rates for the public version, and the performance numbers it has shared—100% detection of 21 planted bugs in a private test driver, 96% recall on five years of clfs.sys cases—come from a controlled environment. In your own codebase, every flagged issue still demands human validation. The AI can flag a flaw and even suggest a fix, but it cannot understand your business logic, deployment schedule, or regression risk.
The commercial product, if it launches this month, will force decisions around licensing. Microsoft has not announced pricing for Project Perception, but The Information’s report notes that security chief Hayete Gallot views the traditional bundling of security features into enterprise agreements with skepticism. Expect a standalone SKU—possibly layered on top of or alongside Defender, GitHub Advanced Security, or Security Copilot. If your team already pays for E5 or Azure services, watch for announcements that could add costs or valuable new capabilities.
For Developers
The GitHub connector is the clearest near-term benefit. Point MDASH at your repo, and it will scan for vulnerabilities before you merge to main. Findings appear in Defender for your security team while also generating work items in GitHub or Azure DevOps. You’ll get code suggestions for fixes, but you still own the build. Test everything in staging. An AI-generated patch that closes a vulnerability while breaking authentication flow or silently degrading performance is a net loss. The loop Microsoft describes—scan, triage, fix, validate, deploy—remains a human-driven process with AI as a high-speed assistant.
How We Got Here: A Swift Ascent
The timeline is startling. In February 2026, Hayete Gallot took over Microsoft’s security organization. Within months, she replaced at least nine corporate vice presidents, laid off several hundred staff, and redirected resources toward AI security teams. That internal upheaval created the bandwidth and focus for MDASH to move from a research project to a customer-facing offering in record time.
Then came the public reveal on May 12, 2026. Microsoft’s security blog detailed how MDASH coordinated over 100 specialized AI agents and a mix of heavyweight and distilled models to hunt bugs. The 16 Windows vulnerabilities it found—and the patches that followed—proved the concept could deliver concrete security improvements, not just benchmark scores.
Meanwhile, the competitive landscape sharpened. Anthropic’s Claude Mythos, a cybersecurity-focused model, has set a high bar for AI-driven vulnerability research. But Mythos access is tightly restricted through Anthropic’s Project Glasswing, limited to vetted critical infrastructure and major software providers. Microsoft sees an opening for a more broadly available service that wraps models inside an operational environment: Defender telemetry, source control, DevOps pipelines, and audit trails. The multi-model approach also insulates the product from overdependence on a single provider’s pricing or availability.
What to Do Now: Immediate Actions
If you’re in the MDASH private preview:
- Integrate with Defender and GitHub now. Connect the scanner to a non-critical repository first. Observe the types and volume of findings before expanding.
- Build a review process. Designate who acts on MDASH findings, what the response SLA should be, and how proposed patches get tested before merging.
- Track false positives and false negatives. Early feedback will shape your trust in the tool. Report inconsistencies through Microsoft’s preview channels.
If you’re not yet in the preview:
- Audit your patching posture. An AI scanner that finds more bugs only helps if you can deploy fixes rapidly. Confirm your update rings, rollback plans, and exceptions list.
- Prepare for Project Perception licensing. If your organization runs on E5, GitHub Advanced Security, or Azure DevOps, start a conversation with your Microsoft account team about how a standalone AI code-security service might sit alongside current spend. Demand early access or a trial.
- Educate your team. AI vulnerability scanners are inevitable. Ensure developers and security staff understand the difference between a scanner that flags and a scanner that patches. Keep humans in the loop.
For everyone:
- Turn on automatic updates if you haven’t already. The faster Microsoft discovers and fixes flaws, the more critical timely patching becomes.
Outlook: A Product Launch and a Patching Reality Check
The next milestone is the reported July launch of Project Perception. That will answer the most pressing commercial questions: price, licensing model, and exactly which models are included under what terms. It will also test whether Microsoft can deliver on the promise of an integrated, multi-model bug-hunting service without overwhelming customers with noise or risky automated fixes.
Longer term, MDASH’s success will be measured not by the number of vulnerabilities it finds but by how quickly those findings turn into safe, deployed patches. The private preview is the first real-world step. The commercial launch will show whether Microsoft can turn an impressive research tool into a product you can trust with your production code.