Codename MDASH: scan code for exploitable vulnerabilities
Security

Codename MDASH: scan code for exploitable vulnerabilities

Codename MDASH uses specialized AI agents to find, validate and prioritize code vulnerabilities for security teams and developers.

Key insights

  • MDASH uses more than 100 specialized agents across preparation, scanning, validation and deduplication.
  • Validation agents debate candidate findings to assess whether they are credible and exploitable.
  • The demo connects source repositories through Defender settings under Settings > Connectors.
  • Scan results can appear in the portal, as an HTML report or as a SARIF file.
  • The presenters described MDASH as being in private preview for some customers.

Codename MDASH scans source code with multiple AI agents to discover, validate and prioritize vulnerabilities, helping teams focus on credible risks and route fixes to developers. It is aimed at security leaders, application security teams and developers who need to assess and remediate flaws in their applications. Microsoft Events demonstrates the workflow in a Microsoft Ignite session, including repository connection, scan review and remediation handoff.

What's new

MDASH is Microsoft Security's multi-model agentic scanning system. Rather than relying on one model or a conventional scanner alone, it coordinates specialized agents and security-focused models to examine code and challenge potential findings. The goal is to distinguish actionable vulnerabilities from candidates that do not hold up to validation.

The system works in four named phases: preparation, scan, validation and deduplication. Preparation ingests code and maps the attack surface; scanning agents propose potential findings; validation agents debate evidence for and against each finding; and deduplication collapses overlapping issues. The result is intended to focus attention on more important findings, including cases where one fix may address multiple reports.

The presenters said the system uses different model types, including frontier reasoning models, faster distilled models and independent counterpoint models. They emphasized that MDASH's approach depends on coordinating models and agent roles rather than on a single model. They also described it as already used inside Microsoft engineering and available to some customers through a private preview program.

Who it is for and what you need

MDASH is relevant to CISOs seeking visibility into application risk, security teams responsible for finding and prioritizing vulnerabilities, and developers who own the code and fixes. The demonstrated workflow connects source code organizations such as GitHub or Azure DevOps, then scans a selected repository. The session's speakers described customer access as a private preview, so teams should treat it as a preview offering rather than assume it is generally available.

In the demo, the security analyst works in Microsoft Defender. The interface includes a source-code connector setup and an MDASH area within Exposure Management. The product view also includes Foundry setup for inference. The example uses a repository named merchant-portal, but it is a demonstration scenario, not a prerequisite or a required repository structure.

How it works

The scan-to-fix workflow moves from connecting repositories to reviewing findings and handing remediation to developers. Microsoft Events' demonstration makes clear that security analysts can identify and prioritize issues without taking over the developers' review and merge process.

  1. In Defender, open Settings > Connectors and connect a source-code organization, such as GitHub or Azure DevOps. The demonstration also shows setting up Foundry for MDASH.
  2. Open Exposure Management, then the MDASH area. Select a connected repository and start a scan. The demo offers standard and MAI Cyber scan profiles.
  3. Review scan status and outputs. The shown status view includes repository information, function counts, AI-validated findings, duration and token use. Results are available in the portal, as an HTML report and as a SARIF file.
  4. In Security Recommendations, filter findings by severity or confidence, then inspect the finding details. The example includes a confidence level, CWE identifier, agent validation discussion and evidence to help reproduce the issue.
  5. Create a fix pull request from the Defender portal when a finding is ready for developer review. Developers can inspect the proposed change in GitHub, review the diff and continue through their normal checks and approval process.

The demo also shows a developer applying fixes with GitHub Copilot and rebuilding the application to validate the result. Those steps illustrate a possible workflow; the presenters stressed that existing organizational gates and developer ownership remain in place. MDASH can support the handoff, while developers retain responsibility for reviewing and moving code changes forward.

Limitations and alternatives

MDASH findings still need to be assessed in the context of the application and its code. The demo exposes evidence and validation details rather than asking an analyst to act on a vulnerability name alone. Reviewing confidence, reproduction context and the proposed remediation helps teams decide whether and how to act.

Agent validation is intended to improve the quality of findings, but it does not replace developer review or organizational testing. In the demonstrated workflow, code changes continue through existing authorization steps and checks before the rebuilt application is treated as fixed. Teams can also use the available HTML or SARIF output rather than relying only on the MDASH portal.

The speakers discussed complex logic issues, including broken object-level authorization, missing authentication and memory-management flaws. The Caldova pharmacy application in the demonstration illustrates how separate authorization problems can combine into operational and customer impact. It is an example of the kinds of issues MDASH is designed to investigate, not a guarantee that every scan will find the same vulnerabilities.

Sources

Frequently asked questions

How does Codename MDASH validate code vulnerabilities?

After scanning agents propose findings, a separate group of agents debates evidence for and against each one. MDASH then deduplicates overlapping findings.

Where do I connect repositories for MDASH?

In the demonstrated Defender workflow, go to Settings > Connectors and connect a source-code organization such as GitHub or Azure DevOps.

What formats can MDASH scan results use?

The demonstrated workflow provides results in the MDASH portal, an HTML report and a SARIF file.

Is Codename MDASH generally available?

The presenters described access for some customers through a private preview program.