Microsoft MDASH Preview Adds AI Vulnerability Scans to Defender CLI

Microsoft’s codename MDASH AI code scanner is now in private preview through Microsoft Security Exposure Management, turning a reported plan to challenge Anthropic’s Claude Mythos into an early enterprise product for finding, validating, and prioritizing software vulnerabilities.
The Information first reported that Microsoft was preparing a tool combining models from Microsoft, OpenAI, and Anthropic to identify bugs and generate fixes. But Microsoft’s July 2026 release notes now show MDASH is available in private preview, accessible through Defender CLI and a GitHub connector, with findings surfaced in the Defender portal. For Windows administrators and security teams, the immediate implication is more practical than a model-versus-model contest: Microsoft is positioning AI-assisted code auditing within the same operational environment already used for exposure management and security triage.
The original Bloomingbit report refers to Anthropic’s model as “Claude Mitos.” The product’s name is Claude Mythos, currently Mythos 5.

MDASH Is a Workflow, Not a Single Model​

Microsoft has been unusually direct about the distinction. In its May security blog post introducing MDASH, the company described the system as a “multi-model agentic scanning harness,” arguing that the product is the surrounding workflow rather than one frontier AI model.
That workflow is designed to move through several stages that traditional static-analysis products often leave to humans: it indexes source code and historical commits, maps attack surfaces, searches candidate code paths, challenges its own findings with separate “debater” agents, removes duplicates, and attempts to prove a bug with triggering inputs where possible. Microsoft says the system employs more than 100 specialized agents.
This matters because raw AI vulnerability reports are not useful if they flood a development organization with speculative bugs. A Windows, Azure, or enterprise software team needs a finding that can be reproduced, assigned to an owner, ranked by impact, fixed, regression-tested, and shipped through a controlled update process. Microsoft’s stated MDASH design aims at that harder problem: reducing the gap between “the model suspects this is unsafe” and “the engineering team can act on this.”
Microsoft says its system combines heavyweight reasoning models with lower-cost models used for broader, repeated analysis passes. The latter act partly as independent reviewers, questioning whether a suspected flaw is reachable or exploitable. That is an important design choice: rather than treating AI agreement as proof, MDASH uses disagreement and challenge as a triage signal.

---

Windows Code Has Already Been the Test Bed​

MDASH is not being presented solely as a lab experiment. Microsoft said in May that the system had found 16 previously undisclosed vulnerabilities across Windows networking and authentication components, including two critical remote-code-execution flaws. The company said those issues were remediated through its security engineering processes.
Microsoft also reported a clean result in a controlled benchmark involving 21 known vulnerabilities in a storage-oriented codebase: all 21 were identified, with no false positives in that particular run. That should be read as a product demonstration rather than a universal guarantee. Benchmarks measure defined cases; production codebases contain build complexity, proprietary libraries, partial dependencies, legacy behaviors, and business logic that can complicate every claimed finding.
Still, Microsoft’s focus on the Windows TCP/IP stack, authentication components, Hyper-V, device drivers, services, Xbox, and Azure is significant. Those are high-value code surfaces where a serious vulnerability can have consequences far beyond a single application. They are also areas where context matters: kernel conventions, interprocess trust boundaries, lock handling, and proprietary interfaces are not easily solved by pattern matching alone.
For Windows users, this does not translate into a client-side MDASH feature or a new Windows Defender toggle. The likely benefit is indirect: better internal and enterprise scanning could move flaws from discovery to Patch Tuesday remediation more quickly, or catch them before products and updates ship.

Microsoft Is Folding AI Scanning Into Defender Operations​

The private-preview placement says as much about Microsoft’s strategy as the underlying model architecture. Rather than launch MDASH as a standalone coding assistant, Microsoft is placing it inside Security Exposure Management. Teams can run scans through Defender CLI or connect GitHub repositories, then review results in the Defender portal alongside their broader exposure picture.
That integration could be more valuable to enterprise customers than a spectacular autonomous bug-finding demo. Security leaders need to know which repositories matter most, which applications have privileged access, which affected assets are exposed, and whether a development team has accepted or remediated the risk. An AI scanner that produces isolated reports is only a new source of work; one connected to vulnerability prioritization and incident workflows has a clearer route to operational use.
It also gives Microsoft a natural channel for a mixed-model product. The Information reported that the new service may use Anthropic, OpenAI, and Microsoft models. Microsoft has not publicly identified every model available in MDASH’s preview, and customers should not assume a particular provider is used for every scan stage. The key public claim is the ensemble approach: models are selected and orchestrated for different roles rather than being asked to independently audit, exploit, explain, and patch a codebase in one pass.
Organizations considering the preview should concentrate on the mechanics that matter:

• Source-code access, build artifacts, secrets handling, and repository permissions will determine whether the product can be deployed safely.
• Human review must remain mandatory before generated proof-of-concept code or patches are executed in production environments.
• Finding quality should be measured against the organization’s existing static analysis, software composition analysis, and secure development lifecycle processes—not judged only by AI-generated severity labels.
• Development teams will need ownership and remediation capacity, because a higher discovery rate can quickly become a patching backlog.

---

Claude Mythos Sets the Competitive Bar—and the Safety Constraint​

Anthropic’s Claude Mythos is the obvious comparison because it has become the industry’s most visible example of an AI system purpose-built for advanced defensive cybersecurity work. Anthropic’s Project Glasswing began in April 2026 with roughly 50 partners, including Microsoft, and has since expanded to about 150 organizations across more than 15 countries.
Anthropic says early partners found more than 10,000 high- or critical-severity flaws across important software systems. Those numbers are vendor-reported and should be interpreted carefully: “found” is not necessarily equivalent to independently confirmed, fixed, or publicly disclosed vulnerabilities. Yet the scale illustrates the operational problem now facing software maintainers. Finding weaknesses is increasingly not the sole bottleneck; verification, coordinated disclosure, patch creation, testing, and deployment are.
Mythos 5 remains restricted to a small group of vetted partners because Anthropic says its cybersecurity and biology capabilities could be misused. Its broader-access counterpart, Claude Fable 5, uses the same underlying model with additional safeguards and routing controls for high-risk requests. Anthropic’s approach underscores the dual-use challenge Microsoft will confront as it expands MDASH: a system that can identify a vulnerability and generate a credible fix may also produce valuable knowledge for an attacker if it is poorly governed.
Microsoft’s answer, at least so far, is to keep MDASH in a private preview and put its code-scanning functions inside a controlled security platform. That is a fundamentally different product posture from handing unrestricted cybersecurity capability to every developer through a general chatbot interface.

The Real Test Will Be Patch Throughput​

Microsoft’s strongest claim is not that MDASH has replaced security researchers. It is that a multi-agent, multi-model system can make expert vulnerability research more repeatable at the scale of Windows and Azure. That is credible as an engineering direction, but the proof will be measured in production outcomes: fewer false positives, faster validation, patches that survive regression testing, and fewer exploitable flaws reaching customers.
The next visible milestone is broader availability and a clearer statement of supported languages, repository requirements, data controls, pricing, and patch-generation safeguards. For now, MDASH’s arrival in Microsoft Security Exposure Management means the AI security race has moved beyond model benchmarks. The companies that matter most to Windows and enterprise IT will be the ones that turn a flood of machine-found flaws into verified fixes before attackers can do the same.

Microsoft is reportedly preparing a commercial AI vulnerability-finding service, internally called Project Perception, that could scan code, validate security flaws and generate fixes using a mix of Microsoft, OpenAI and Anthropic models. But Windows administrators should separate the report’s prospective product from what is already available: Microsoft’s own documentation says its codename MDASH agentic code scanner entered private preview in Microsoft Security Exposure Management this month.
The Information, which first reported the Project Perception plans on July 16, says Microsoft could debut the product as soon as July and intends it as a lower-cost alternative to Anthropic’s restricted Claude Mythos offering. The reported product would go beyond a conventional static-analysis scanner, coordinating different models for bug discovery, root-cause analysis and patch generation.
That is a consequential shift for Windows, Azure and enterprise development teams. Microsoft is no longer presenting AI-assisted vulnerability research as an internal experiment; it is beginning to wire the capability into the Defender portal, GitHub connectors and remediation workflows that customers already use.

MDASH Is the Technical Foundation, Not a Magic Patch Button​

Microsoft introduced its Multi-Model Agentic Scanning Harness, or MDASH, in May 2026. The company describes it as a vulnerability discovery and remediation system built around more than 100 specialized agents and a panel of heavyweight and distilled AI models, rather than one all-purpose model.
Microsoft’s security blog said MDASH helped researchers uncover 16 previously unknown vulnerabilities in Windows networking and authentication components. Four were critical remote-code-execution issues, including flaws involving the Windows kernel TCP/IP stack and IKEv2. The May Patch Tuesday release included fixes for that group of findings, tying the system to a real Windows servicing outcome rather than a benchmark-only demonstration.
The company’s public performance claims are impressive but deserve the usual vendor caution. Microsoft says MDASH found all 21 planted bugs in one private test driver without false positives, reached 96% recall across five years of confirmed MSRC cases in clfs.sys, and scored 88.45% on the public CyberGym benchmark. Those figures show why Microsoft sees a market opportunity, but they do not mean an organization should allow AI-generated patches into production unattended.
For its current private preview, Microsoft says teams can run MDASH scans from Defender CLI or a GitHub connector, review findings in the Defender portal, and use them to prioritize code-security risk. That wording matters. It describes a human-led security workflow augmented by AI, not automatic changes being merged into Windows images, application repositories or production infrastructure.

---

The Product Story Is Bigger Than a Scanner​

The Information’s report adds a commercial layer to MDASH. Project Perception is said to use a model router that picks among Microsoft, OpenAI and Anthropic models depending on the task. The logic is straightforward: expensive reasoning models can be reserved for difficult validation and exploitability questions, while less costly models handle high-volume scanning, duplicate detection and preliminary triage.
That multi-model design is arguably the more important news than the reported Project Perception name. It avoids betting an entire security product on a single model provider’s capabilities, pricing, availability or safety restrictions. It also gives Microsoft an avenue to treat vulnerability research as an orchestration problem: prove the defect, determine whether it is reachable and exploitable, produce a focused fix, and test whether the change breaks something else.
Microsoft has already described that workflow in broad terms. MDASH findings can flow into Microsoft Defender for prioritization alongside runtime signals and threat intelligence, then into GitHub and Azure DevOps for work-item creation, pull requests, validation and remediation. For organizations that run Windows-heavy estates alongside GitHub Enterprise, Azure DevOps and Defender, that closed loop is the potential differentiator.
The obvious challenge is trust. Automated code changes can fix a vulnerability while creating a regression, reducing performance, weakening compatibility or simply masking the original condition. Microsoft has repeatedly said human researchers and engineering teams remain in the process, and that is exactly where mature security organizations should keep them. AI can make the funnel wider; it cannot eliminate the need for code owners, test suites, release controls and rollback plans.

Anthropic’s Mythos Sets a Different Kind of Benchmark​

The comparison target is Claude Mythos, not “Claude Mitos.” Anthropic’s cybersecurity-focused model is designed for advanced vulnerability research and exploit reasoning, but access is tightly controlled through its Project Glasswing program. Anthropic says Mythos 5 is available only to a limited set of vetted partners because the same capabilities that help defenders find flaws can also be misused to develop attacks.
That policy creates an opening for Microsoft, but it also sets a high bar. Mythos is not merely marketed as a code-review assistant; Anthropic positions it as a frontier capability intended for critical infrastructure and major software providers. Anthropic’s broadly available Claude Fable 5 uses the same underlying model family with additional cyber and bio safeguards, while the more capable Mythos path remains gated.
Microsoft’s prospective offering appears to pursue a different balance. Rather than distributing one especially capable cybersecurity model, Project Perception reportedly would package multiple models inside an enterprise product with a security workflow around them. The practical distinction is not academic: Anthropic is controlling model access, whereas Microsoft’s opportunity lies in controlling the operational environment—identity, source repositories, Defender telemetry, DevOps pipelines, audit trails and remediation approval.
For IT departments, the relevant question will not be whether Microsoft has a more capable chatbot. It will be whether the service can reduce time-to-triage and time-to-fix without flooding security and engineering teams with speculative findings or risky patches.

---

Gallot’s Security Reorganization Gives the Launch Its Context​

The Information reports that Microsoft security chief Hayete Gallot has made AI security a central priority since taking over the organization in February 2026. According to the publication, at least nine corporate vice presidents who previously reported into the security organization have departed this year, while teams focused on older products have faced cuts and AI-oriented groups have expanded.
The reported restructuring also suggests Microsoft may revisit a long-standing enterprise sales habit: bundling security features into broader Microsoft agreements. If AI-powered code security becomes a major standalone offering, pricing and licensing will matter as much as the models underneath it. The Information said Project Perception pricing had not been finalized.
That is particularly relevant for Windows and Microsoft 365 customers already paying for E5, Defender products, GitHub Advanced Security, Security Copilot or Azure services. A new AI code-security service could arrive as a premium add-on, an Exposure Management capability, a Security Copilot consumption feature, or some combination of those. Microsoft has not publicly confirmed the Project Perception branding, release date, price or licensing model.

More Findings Mean a Harder Patching Discipline​

The near-term effect of MDASH may be less glamorous than autonomous patching: more vulnerabilities will be found and disclosed. Microsoft has already put the system to work across Windows, Azure and identity engineering, targeting complex code where manual review is expensive and incomplete.
That is good news only if organizations can absorb the output. Security teams will need better asset inventories, dependable patch rings, evidence-based prioritization and a clear process for exceptional systems that cannot be updated immediately. An AI system that finds five times as many plausible flaws does not automatically make an enterprise five times safer if remediation queues, change-management windows and application testing remain unchanged.
Microsoft’s private preview is the immediate milestone to watch. A broader Project Perception launch, if it arrives in July as The Information reported, will show whether the company can turn MDASH from an internal Windows bug-finding engine into a product that helps customers close the same gap between discovery and remediation.