AI
GitHub and Microsoft pilot AI that fixes vulnerabilities, not just finds them
A six-project experiment delivered verified security patches to open source maintainers, cutting triage time from hours to minutes.
Key takeaways
- GitHub and Microsoft sent 99 verified vulnerability reports with suggested fixes to six open source projects; maintainers fixed 55 in weeks, some in under 15 minutes.
- Codename MDASH used AI to scan, an AI agent to triage, and seven human researchers to verify every finding before delivery—no raw scanner output reached maintainers.
- Maintainers said the approach surfaced obscure bugs and cut triage time dramatically compared to typical AI-generated security reports.
- GitHub will offer verified AI security scans to all alumni of its Secure Open Source Fund and is building the workflow into private vulnerability reporting.
- The pilot exposed gaps: maintainers need private CI/CD testing for fixes and access to cyber-capable models to verify exploits before and after patching.
Open source maintainers are drowning in AI-generated security reports. Many are duplicates, false positives, or findings that don't account for a project's actual threat model. The real bottleneck isn't discovering vulnerabilities anymore—it's fixing them.
GitHub and Microsoft ran a pilot to flip that dynamic. Instead of flooding maintainers with raw scanner output, they used an AI system called Codename MDASH to find vulnerabilities, filter them through automated triage, verify them with human researchers, and deliver only high-confidence reports with suggested fixes already attached.
How the process worked
The pilot involved six open source projects, all alumni of the GitHub Secure Open Source Fund: AutoGPT, Caddy, Express, FastMCP, ImageMagick, and Netty. Each maintainer consented to the scan upfront.
Codename MDASH, built by Microsoft Security and powered by the MAI Cyber model, scanned each project and flagged thousands of potential issues. An AI triage agent from GitHub Security Lab—built on Copilot skills—removed duplicates and filtered for validity and impact. Seven researchers from Microsoft FORGE Lab then manually verified every remaining finding.
Only after that three-layer review did maintainers receive reports, delivered privately through GitHub's vulnerability reporting system. Each report included a proof of concept and a suggested fix.
The result: 99 private reports sent. As of mid-November, maintainers had fixed 55 issues, with 14 already released. Caddy's team resolved all 22 reports in under a week. One ImageMagick fix took 15 minutes, down from the team's typical 30 to 60 minutes per advisory.
Why it matters
This approach shifts the burden of triage and verification away from maintainers and onto the organizations with resources to do it. For a two-person team like ImageMagick's, that difference is material. For larger projects, it cuts through noise that obscures real issues.
The pilot also exposes infrastructure gaps. Private vulnerability reporting needs broader adoption. Maintainers need CI/CD workflows that can test fixes on private forks. And defenders need access to the same cyber-capable models attackers already use.
What maintainers said
Express maintainers said the reports cut through noise. Caddy's team noted the scan surfaced obscure bugs missed by earlier tools. FastMCP maintainers called the findings more useful than previous AI scanner output. AutoGPT received a report three days after shipping a broken security check; the team used it to write a better fix and had it in production within a week.
One Netty maintainer put it plainly: "Finding vulnerabilities is no longer our bottleneck. Fixing them is the hard part."
What's next
GitHub is integrating these lessons into its private vulnerability reporting and security advisory workflows. Microsoft announced MAI Cyber 2, the next version of the model used in the pilot. GitHub will offer AI security scans—with findings verified before delivery—to all Secure Open Source Fund alumni, plus GitHub Copilot Max access.
The standard GitHub and Microsoft are proposing: check every finding against the project's existing knowledge and threat model, report privately, and include a suggested fix. Let maintainers control the pace.
The details were first reported by GitHub in a post on the GitHub Blog.