Blog Post

Microsoft Defender for Cloud Blog
7 MIN READ

Built to Protect: The Architecture Behind Codename MDASH

JasonWeber's avatar
JasonWeber
Icon for Microsoft rankMicrosoft
Jul 30, 2026

How Microsoft Security is responsibly bringing agentic vulnerability discovery to customers with Defender class security, sovereignty, compliance, and operational governance.

Information Table

Defender
Foundations

Regional
Operation

AI & Data
Containment

Responsible
Governance

Built as a production Microsoft Security service, not an isolated research system.

Operates alongside Microsoft Defender in the clouds and regions customers already use.

Designed to keep source, findings, and security data inside controlled boundaries.

Security, privacy, compliance, responsible AI, and operational readiness.

A new class of AI security needs a new level of trust

When Microsoft introduced Codename MDASH, we shared a major step forward in AI-powered vulnerability discovery: a multi-model, agentic scanning system that helps customers discover, validate, and reason about exploitable vulnerabilities across  software applications. MDASH brings together specialized agents, multiple models, domain-specific analysis, debate, and proof construction to move beyond theoretical alerts and toward findings that customers can remediate and act on.

That capability matters because the software landscape is changing quickly. AI is accelerating how code is written, how systems are analyzed, and how vulnerabilities are discovered. The same advances that help defenders find critical weaknesses earlier can also lower the barrier for attackers and make sophisticated vulnerability research more accessible. In that environment, customers need more than powerful AI. They need powerful AI delivered inside a trusted security architecture.

That is the design behind MDASH. The product is not simply the model and harness. The product is the comprehensive security system that surrounds them: the controls, isolation boundaries, operational practices, governance, and compliance posture that allow customers to benefit from frontier AI and advanced vulnerability discovery without compromising control, and leveraging the discoveries inside Microsoft Defender.

Built as a production Microsoft Defender service

MDASH was built from the beginning as a production Microsoft Security service. It was not designed as a standalone experiment, a separate research environment, or a disconnected security tool. It was designed to operate with the same security, governance, privacy, and operational foundations customers expect from Microsoft Defender XDR and Microsoft Sentinel.

That distinction is important. Security products operate on some of the most sensitive information an organization has: source code, vulnerability findings, exploitability analysis, threat intelligence, identity signals, runtime exposure, security incidents, and remediation workflows. Customers should not have to move that data into a new trust model simply to benefit from AI. MDASH is designed to bring AI-powered vulnerability discovery into the trusted Microsoft Defender ecosystem instead.

By operating as part of Microsoft Security, MDASH benefits from Defender class operational discipline including identity and access control, service-to-service authorization, logging and auditability, security review gates, compliance alignment, privacy review, responsible AI review, and production operations. The goal is for customers to experience MDASH as an extension of the Microsoft security platform they already trust, not as a separate AI system they need to evaluate from first principles.

Operating globally where Microsoft Defender operates

Security decisions today are increasingly evaluated alongside data governance, regulatory, and operational requirements. Customers want to know where their data is stored, which services process it, who can access it, and how operational access is governed. These requirements are especially important for AI security systems because the data being analyzed may include proprietary source code, high-severity vulnerabilities, unreleased security findings, and sensitive threat context.

MDASH was designed with that reality in mind. It operates alongside Microsoft Defender in the clouds and regions where Defender operates, allowing customers to benefit from regional deployment models, data residency commitments, sovereign cloud investments, customer governance capabilities, and Microsoft’s broader compliance commitments. The architecture is designed so that vulnerability discovery does not become a reason for customers to move sensitive security data outside the boundaries they rely on today. MDASH integrates with Microsoft Foundry for AI inference. Data processed through Foundry remains subject to the applicable Foundry data handling and compliance commitments.

 

When customers use MDASH to analyze their code, the resulting security context is kept within the customer’s broader Defender security data. Findings, repository metadata, scan artifacts, vulnerability analysis, and security intelligence are treated as part of the customer’s Defender security context. That enables a unified experience while preserving the controls customers require.

Designed so data stays secure

The most important architectural principle behind MDASH is containment. AI systems that work with security data must be treated differently from general-purpose productivity systems. A vulnerability report can be a roadmap for an attacker. Source code can contain intellectual property and secrets. Threat intelligence can reveal what defenders know and how they detect adversaries. Customer security posture data can expose where an organization is weak.

MDASH was built around the principle that the AI should operate like a highly privileged security researcher inside a controlled Microsoft Security environment: powerful enough to do useful work, but isolated, monitored, constrained, and governed so that customer data cannot escape the trusted boundary.

Customer repositories are processed in isolated sandbox environments with dedicated execution boundaries. Scans are separated to prevent persistence, cross-tenant access, or interference between workloads. Agents are provisioned for the job, execute within constrained conditions, and are cleaned up when the work is complete. The architecture is designed to reduce the lifetime of sensitive data and require a minimum scope to perform the analysis.

Intelligence belongs to the organization

Security data is valuable, but so is the intelligence generated while analyzing it. MDASH was built around the principle that AI should help organizations leverage intelligence without requiring them to surrender it. The goal is to ensure that organizations can benefit from advanced AI capabilities without introducing new risks to their intellectual property, operational knowledge, or security practices.

MDASH is designed so that context remains local to the customer's tenant. Intelligence generated during scanning and investigation is not treated as exhaust to be harvested, but as an asset that belongs to the organization. This architectural approach helps organizations adopt AI while maintaining ownership of their security expertise, preserving competitive advantage, and ensuring that the value created through their interactions remains inside the trusted boundary where it was generated.

Secure isolation as a first-class control

Network behavior is one of the most important risk areas for any AI harness. Many open-source and experimental harnesses assume broad network access. Agents may call external websites, package registries, public APIs, malware analysis services, paste sites, or search engines as part of their workflow. For defensive security workloads, that model presents risks. It can create paths for source code, findings, prompts, exploit details, or threat intelligence to leave the environment.

MDASH was designed differently. Critical services are placed behind private networking boundaries. Communications occur through controlled service paths. Ingress and egress are restricted, monitored, authorized, and validated. Service-to-service communication is explicit. External communication is controlled.

Similarly, the environment is closed to new code and dependencies. The AI cannot download tools, install packages, import attacker frameworks, or otherwise expand at runtime. By preventing the introduction of untrusted software, we eliminate one of the most common patterns seen in AI escape and control-bypass research, where a model gains access to new tooling that enables actions beyond its intended boundaries.

This matters for both sides of the trust boundary. Customers need confidence that their source code, vulnerability findings, and security data are not exfiltrated to unauthorized destinations. Microsoft also needs confidence that its own threat intelligence, security research, and detection context can be used to improve analysis without leaking outside the protected environment. Isolation is a core product promise.

Tenant isolation, identity, and authorization

MDASH is designed for multi-tenant enterprise use, and tenant isolation is one of the most heavily scrutinized parts of the architecture. The platform is built around the principle that one customer’s code, findings, scan history, telemetry, or generated security intelligence must never become visible to another customer.

That principle is enforced through multiple layers including identity validation, tenant-aware authorization, role-based permissions, service-to-service authentication, storage isolation, queue isolation, telemetry separation, and audit logging. No single control is expected to carry the entire security model. Instead, the architecture layers controls so that customer data remains separated across the API, processing, storage, telemetry, and operational planes.

For customers, the practical outcome is straightforward. MDASH can operate on highly sensitive source code and security findings while preserving the same separation, access discipline, and audit expectations customers require from Microsoft Defender.

Encryption, retention, and data protection

Customer source code repositories and findings are protected through encryption in transit and at rest. Secrets are handled through enterprise secret management systems. Access to customer data is governed through authorization controls and operational processes. The platform is designed to keep sensitive data scoped to the operation being performed and retained only as long as necessary to provide the service.

Data minimization is especially important for AI security systems. The goal is not to accumulate customer code or retain vulnerability artifacts indefinitely. The goal is to perform the analysis, generate the security value, connect the finding to Defender context, and preserve only what is needed for the customer’s security workflow, auditability, and remediation experience.

This is also why MDASH is designed around a clear separation between customer data and model improvement. Customers need confidence that their proprietary code, vulnerabilities, and security findings are not being used in ways they did not authorize. Trust requires clear boundaries.

Responsible AI, security review, and operational governance

Responsible AI is not a separate layer added at the end of a security product. For MDASH, it is part of the engineering model. The system is powerful precisely because it can reason about exploitable vulnerabilities; that same power requires careful governance, human oversight, transparency, and operational controls.

MDASH has been developed with extensive review across security, privacy, compliance, Responsible AI, architecture, authorization, tenant isolation, penetration testing, and operational readiness. These reviews are not ceremonial. They are how Microsoft ensures that advanced AI capabilities are introduced with the discipline customers expect from enterprise security products.

The objective is not only to build an AI system that can find vulnerabilities. The objective is to build a trustworthy security service that customers can deploy in regulated, sensitive, and globally distributed environments. That is the standard Microsoft believes this category requires.

Updated Jul 29, 2026
Version 1.0

2 Comments

  • nawarrajohn's avatar
    nawarrajohn
    Copper Contributor

    Thank you for this article.  While there are explanations that seem to imply it, it would have been nice to read the statement, "Microsoft will not train its AI models using the customer's code repository", or even better, augmented with "... nor any other information obtained from the customer during MDASH usage".  

  • Dean_Gross's avatar
    Dean_Gross
    Silver Contributor

    this is very interesting, but I was hoping to see some diagrams and technical information about the actual architecture.