What matters in AI.

Subscribe

GitHub and Microsoft send open source teams 99 checked bug reports

As of 15 November, maintainers have fixes for 55 of the 99 private reports, and 14 fixes are in a release.

GitHub and Microsoft test Codename MDASH, a Microsoft system that uses some AI models to examine code for security weak points. It flags thousands of possible issues in six open source projects that many teams use. AI triage and seven Microsoft researchers decrease them to 99 private reports. Each report has a proof of concept and a fix.

How it works

The test has these steps:

  • Codename MDASH examines the code and gives fixes.
  • A GitHub AI agent removes duplicates and weak results.
  • Microsoft researchers examine each remaining result by hand.
  • Maintainers get a private report and select the changes and the time.

The numbers

Caddy maintainers make fixes for all 22 of their reports. An ImageMagick maintainer tells that a report usually uses 30 minutes to 1 hour, and this one uses 15 minutes. Researchers use about two days for each project.

Unknown points

Six projects is a small sample, and the lessons come from what the maintainers share. The hand checks cannot be done for more projects at this time. Reports of low quality continue to go to maintainers each day. The good comments come from partners in the test.

Insight: Checks by persons cost more than the search for weak points

In October, Anthropic starts a free scanner. Its reports have no check by a person, and GitHub AI Scan results cannot stop a merge. This test adds hand checks and fixes, at about two days of work for each project. It is easy to find weak points, and a check by a person is the part that is not easy. A vendor that sends checked fixes seems to get more use from maintainers.

Sources

  1. Maintainers don’t need more AI bug reports. They need fixes.github.blog
AI MATTER · NEWS · AI MATTER · NEWS ·17 NOV2026

Posted

Tags

More in Agents

All Agents news