AlphaGazette
Menu

Security

Open-source maintainers fixed 55 of 99 AI-vetted security reports

A GitHub and Microsoft pilot used Codename MDASH and Microsoft’s MAI Cyber model, then human review, before any report reached a maintainer. As of November 15, 14 of the fixes were already in a release.

1 min read AI-generated

GitHub and Microsoft say a joint pilot turned AI vulnerability scans into patches that open-source maintainers shipped. On six projects, maintainers had closed 55 of 99 private reports as of November 15, and 14 of those repairs were already out in a release.

Codename MDASH, a multi-model agent harness from Microsoft Security, scanned each codebase with Microsoft’s MAI Cyber model and drafted a repair for every issue it raised. That first pass marked thousands of candidate problems. A triage agent from the GitHub Security Lab, built with customized Copilot skills, then dropped repeats, tested whether a finding was real, and kept the bugs with greater impact, so maintainers never saw the raw scan. Seven researchers at Microsoft FORGE Lab reviewed whatever remained. Only then did a report go out on GitHub’s private channel for vulnerabilities, with a proof of concept and a proposed patch. Maintainers still chose the change, the method, and the timing, and both the reports and the patches stayed on GitHub.

All six participants are alumni of the GitHub Secure Open Source Fund. They had agreed to the scan and received every report in private. One of them is AutoGPT, a platform for AI agents. GitHub and Microsoft set two rules for the work: forward only issues that matter, and hold back any finding a person has not checked.

GitHub said inexpensive AI scanning has left many maintainers buried in machine-written security notes, including duplicates and notes that are not flaws. Log4j maintainer Christian Grobmeier has argued that AI used to protect open source must beat AI used to attack it. GitHub’s advice for anyone aiming a new model at a project is to ask the maintainers first and to show up with patches.