On November 17, 2026, GitHub and Microsoft published results from a pilot called Codename MDASH, a multi-model AI system that scanned six widely used open-source projects for security vulnerabilities — and, unlike most AI bug-scanning tools, didn’t just dump the results on maintainers. The pipeline filtered roughly 9,791 raw AI findings down to 99 verified, private reports, each shipped with a suggested fix. As of November 15, maintainers had already fixed 55 of them.

The problem MDASH was built to fix

AI scanners have gotten good at finding potential vulnerabilities in code. The trouble, as GitHub put it in its own announcement, is that “maintainers don’t need more AI bug reports — they need fixes.” A scanner that floods an unpaid maintainer’s inbox with thousands of unverified “possible issues” doesn’t save anyone time; it just moves the triage burden onto the person least equipped to absorb it. Christian Grobmeier, a maintainer of the Java logging library Log4j — the project at the center of the 2021 Log4Shell vulnerability — framed the stakes bluntly: “AI defending open source has to be better than the AI attacking it.”

How the pipeline actually works

Codename MDASH runs in five stages. First, Microsoft’s MAI Cyber model scans a project and proposes fixes. Second, an AI triage agent built by GitHub’s Security Lab — using tailored Copilot skills — removes duplicates, checks validity, and narrows the list to high-impact bugs. Third, human researchers from Microsoft’s FORGE Lab manually check every remaining finding; GitHub says seven researchers spent roughly two days per project doing this by hand. Fourth, verified findings go to maintainers privately through GitHub’s existing private vulnerability reporting system, each bundled with a proof of concept and a suggested fix. Fifth — and this is the part GitHub is explicit about — the maintainer still decides what to fix, how, and when. The AI doesn’t commit code on its own.

That human-verification step is what separates the 9,791-to-99 ratio from a typical AI scanner’s output. GitHub frames it as roughly one checked report for every 100 raw findings — the opposite of the usual AI-security complaint, where false positives outnumber real bugs.

Six real projects, six real outcomes

The pilot covered six alumni of GitHub’s Secure Open Source Fund: AutoGPT (an AI agent platform), Caddy (a web server), Express (Node.js’s dominant web framework), FastMCP (a framework for building Model Context Protocol servers), ImageMagick (the long-running image-processing suite), and Netty (a Java networking framework). The results varied by project size and staffing. Caddy’s maintainers received 22 reports and fixed every one within a week. ImageMagick — run by a two-person team — also received 22 reports and closed one in 15 minutes. AutoGPT got a report on a broken, recently shipped check three days after the scan ran, and had a fix in production less than a week later. Netty’s maintainers put the shift in priorities plainly: “Finding vulnerabilities is no longer our bottleneck. Fixing them is the hard part, and that depends on maintainers.”

What happens next

GitHub says it will extend AI security scans — with the same human-checked-before-reported model — to every alumni project of the Secure Open Source Fund, along with access to GitHub Copilot Max, and will carry the support into future Fund sessions. The Fund itself has now backed 216 projects with more than 300 maintainers, supporting software GitHub says is downloaded over 20 billion times a month; each project in the program receives $10,000 through GitHub Sponsors plus security training and Azure credits from Microsoft for Startups. Notably, the MDASH pilot ran on Microsoft’s earlier MAI Cyber model — the same day GitHub published these results, Microsoft separately announced its successor, MAI Cyber 2, which has not yet been used in this program. GitHub has not said whether Copilot Max access for alumni comes at a cost or how often future scans will run, so maintainers curious about joining the next round should watch GitHub’s Secure Open Source Fund page rather than assume terms from this pilot.

Image: illustrative stock photo of code on a screen; not an actual screenshot of the MDASH tool, which has not been made publicly viewable. Photo by Chris Ried / Unsplash.