GitHub and Microsoft's MDASH Pilot: AI Scanning Yields 99 Verified Private Reports, With 55 Fixed Across 6 Open-Source Projects
GitHub says the Codename MDASH pilot with Microsoft narrowed thousands of raw AI findings to 99 private reports with suggested fixes for 6 open-source projects. As of Nov. 15, 55 were fixed. The sample is small, so results are early.
News date: About 6 min read
Contents
In a post on November 17, 2026, GitHub said it ran the Codename MDASH pilot with Microsoft on 6 open-source projects. The point was not to send more AI-generated vulnerability reports, but to send only verified private reports with suggested fixes. This is an account of a pilot's results, not a service open to all open-source projects; everything below is as stated by GitHub and Microsoft.
What did the pilot do, and how does it differ from ordinary AI scanning?
GitHub notes that AI has made finding potential vulnerabilities fast and cheap, but fixing them still takes human judgment, and many maintainers are swamped by AI reports that are duplicates or not vulnerabilities at all. The pilot aimed to send only the findings that matter and never to let maintainers receive unverified ones. The process had five steps:
- Find: Microsoft Security's Codename MDASH scanned each project using Microsoft's MAI Cyber model and suggested fixes.
- Filter: an AI triage agent built by GitHub Security Lab with Copilot skills (triage means deciding which findings deserve attention) removed duplicates, checked validity and focused on high-impact issues, so maintainers did not have to see raw scan output.
- Verify: Microsoft FORGE Lab researchers manually checked every remaining finding.
- Report privately: findings were sent to maintainers through GitHub private vulnerability reporting, with a proof of concept (code or steps showing the vulnerability is real) and a suggested fix.
- Fix: maintainers decided what to fix, how and when.
In its May 12 blog post, Microsoft described MDASH as a multi-model agentic scanning harness. Agentic means it is built from AI agents that carry out tasks step by step. The harness coordinates more than 100 specialized AI agents (AI programs each handling a specific task), with a pipeline of preparation, scanning, validation, deduplication and proving stages. Microsoft said the system is used by its internal security engineering teams and is being tested by a small number of customers in a limited private preview.
What were the results, and what are the numbers and limits?
| Item | What GitHub reported |
|---|---|
| Pilot projects | AutoGPT, Caddy, Express, FastMCP, ImageMagick, Netty |
| Final private reports | 99 (filtered from thousands of possible issues) |
| Fixed (as of Nov. 15) | 55, of which 14 were already in a release |
| Manual review effort | 7 researchers, about 2 days per project |
| Caddy | 22 received, all fixed within a week |
| ImageMagick | Two-person team received 22; for one report, maintainers pushed a fix in 15 minutes |
| Model used | Pilot used the earlier MAI Cyber model; MAI Cyber 2 was announced by Microsoft the same day |
GitHub said the earliest fixes shipped about a week after the first scan. In a quote published by GitHub, the ImageMagick maintainer said a typical security advisory takes 30 minutes to 1 hour on average to read, understand and fix. GitHub also stresses that 6 projects is a small sample and the results should be treated as early lessons.
What should maintainers and people submitting AI reports do?
GitHub recommends that maintainers enable private vulnerability reporting so researchers can report security issues without disclosing them publicly; for projects on GitHub, gh-secure can turn on core security features in about two minutes. People who depend on open-source software can enable Dependabot alerts to be notified when dependencies they use need security updates. Repositories using GitHub code security can use agentic autofix with Copilot to help fix alerts automatically.
For people submitting AI-generated reports to open-source projects, GitHub relays the standard drawn from maintainer feedback: check each finding against what the project already knows, its threat model and its security policy; report privately; include a suggested fix; and let the maintainer set the pace.
What happens next?
GitHub says it will bring these lessons into private vulnerability reporting and security advisories. It also says it will offer AI security scanning to every alumni project of the GitHub Secure Open Source Fund, with findings checked first and then reported privately. Alumni also get GitHub Copilot Max, and GitHub will bring this support to future sessions of the fund. GitHub says the fund has so far covered 216 projects, with more than 300 maintainers fixing more than 4,000 vulnerabilities.
Frequently asked questions
Is MDASH available to everyone now?
Based on Microsoft's description, no. In May, Microsoft said MDASH is used by internal teams and tested by a small number of customers in a limited private preview; GitHub's post describes a pilot with 6 projects.
Why include suggested fixes and use private reporting?
GitHub says fixing is where maintainers spend the most time, so reports include checked suggested fixes that maintainers can review, adjust and merge. Private reporting lets researchers report issues without disclosing them publicly.
What do users of these projects need to do?
GitHub gives no separate instructions for ordinary users. It only suggests that people who depend on open-source software enable Dependabot alerts to learn when dependencies they use need security updates.
Can these results be generalized to all open-source projects?
No. GitHub itself says the sample is only 6 projects, and manual review took about two days per project and does not yet scale, so the results should be treated as early lessons.
Browse the latest news in this topic
Found this guide helpful? Share it with someone who needs it
NVIDIA and Microsoft at Ignite: Vera Rubin NVL72 in Preview on Azure, Nemotron 3.5 in Foundry, Windows 365 Direct Stream in Limited Preview
NVIDIA says it has expanded AI collaboration with Microsoft from the Azure cloud to the edge and Windows, including a Vera Rubin NVL72 preview on Azure, Nemotron 3.5 models in Foundry and a limited preview of Windows 365 Direct Stream.
Google Quantum AI Makes the Cover of Science: Quantum Contextuality Holds at Up to 105 Qubits, With a 98.3% Win Rate in a Coordination Game
Google: ARMORY CASTLE Exploited VMware vCenter Zero-Day CVE-2026-123890, Affecting vCenter 7.x and 8.x; Broadcom Has Released a Patch
Cloudflare: Election-Day Traffic to 2026 U.S. Midterm Election Sites Surged 363%, While Attackers Shifted from DDoS to Exploiting Vulnerabilities
Sources
- Maintainers don't need more AI bug reports. They need fixes. - The GitHub BlogChecked:
- Defense at AI speed: Microsoft’s new multi-model agentic security system tops leading industry benchmark | Microsoft Security BlogChecked:
- Microsoft for Startups | MicrosoftChecked:
- Microsoft launches new European Security Program - Microsoft On the IssuesChecked:
- Secure Future Initiative – Secure by Design | MicrosoftChecked:
Compiled from the references and verification dates listed above; fares, schedules, product features, and terms can change, so please confirm the latest details on the relevant official website before travelling, booking, or purchasing.