The development

What happened

GitHub says a six-project pilot combined Microsoft's MAI Cyber model in Codename MDASH with AI triage and human researcher checks. Thousands of potential security issues became 99 private reports with suggested fixes. By 15 November, maintainers had fixed 55; 14 fixes were in released software.

Key facts

  • The participating projects were AutoGPT, Caddy, Express, FastMCP, ImageMagick and Netty. Maintainers agreed to scanning and controlled decisions about fixes.

  • Seven Microsoft FORGE Lab researchers checked findings by hand, spending about two days per project. GitHub says that careful process does not yet scale.

  • Reports were sent privately with a proof of concept and suggested fix, rather than passing raw AI findings to maintainers.

  • The pilot used the earlier MAI Cyber model, not the newly announced MAI Cyber 2.

The context

Why it matters

TideReach assessment: The useful result is a checked path from AI findings to maintainer-owned fixes, not an autonomous success rate. Six selected projects and substantial human review do not establish accuracy or staffing needs across open source; fixed and released counts remain different outcomes.