The Pulse
GitHub and Microsoft pilot turns AI findings into 99 checked reports
A pilot sent 99 private vulnerability reports to six open-source projects after AI triage and human review. Maintainers had fixed 55 of the reported issues as of November 15, GitHub says.

A GitHub and Microsoft pilot turned thousands of possible AI-flagged security issues into 99 checked, private reports for six open-source projects, according to GitHub. As of November 15, maintainers had fixed 55 of the reported issues, with 14 of those fixes already in a release.
Source github.blog
The pilot used Codename MDASH, a Microsoft Security system, to scan projects with Microsoft's MAI Cyber model and suggest fixes. The projects were AutoGPT, Caddy, Express, FastMCP, ImageMagick and Netty.
From scan to private report
GitHub says maintainers have been receiving a flood of AI-generated security findings, including duplicates and reports that do not describe vulnerabilities. The pilot was designed to keep raw scan results away from maintainers and provide suggested fixes alongside checked reports.
An AI triage agent from GitHub Security Lab, built on tailored Copilot skills, removed duplicates, checked validity and focused on high-impact bugs. Seven researchers from Microsoft FORGE Lab then reviewed the remaining findings by hand.
The resulting reports went to maintainers through GitHub private vulnerability reporting, each with a proof of concept and a suggested fix. Maintainers decided whether and how to fix issues and when to do so; reports and fixes stayed on GitHub.
What maintainers fixed
The first fixes shipped about a week after the initial scan, GitHub says. Caddy maintainers fixed all 22 reports they received in a week or less. Some suggested fixes were used as written, while others were changed.
AutoGPT had recently shipped a new check that was broken, according to the company. The report arrived three days later and pointed to the relevant line; maintainers went beyond the suggested patch, wrote a fix within a day and had it in production less than a week later.
ImageMagick, maintained by a two-person team, received 22 reports. For one issue, the report's summary pointed to the relevant lines of code, and the maintainers pushed a fix in 15 minutes, GitHub says. An Express maintainer said, “Codename MDASH is very useful for cutting through the noise of vulnerability reports.”
The cost of checking
The company cautioned that six projects were a small sample and described the results as early lessons. The seven researchers spent about two days per project checking findings by hand, and GitHub said that work does not yet scale.
GitHub recommends that people sending AI-generated vulnerability reports check findings against a project's existing knowledge, threat model and security policy, report them privately and include a suggested fix. It also says maintainers need to be able to test fixes on private forks and use cyber-capable models to verify proofs of concept before and after a fix.
Plans for fund alumni
GitHub and Microsoft are working to scale researcher review, and GitHub plans to offer checked AI security scans to every alumni project of the GitHub Secure Open Source Fund. Findings will be checked before being reported privately; alumni also receive GitHub Copilot Max.
All six pilot projects are fund alumni. Each project receives $10,000 through GitHub Sponsors, plus security training and access to an alumni community; GitHub says it will bring the scan support to future fund sessions.
Microsoft announced MAI Cyber 2 on November 17. The pilot used the earlier MAI Cyber model.
Source
Companies in this story
Related stories
- Vercel opens paid Security Review to unlinked GitHub repositories
- CleanStart report flags risks in AI-recommended software dependencies
- Microsoft researchers show neuron edits can backdoor AI agents
- MatrAIx opens proposals for long-horizon personal agent benchmark
- Cloudflare open-sources experimental Kitesurf browser for AI agents