GitHub and Microsoft Pilot Turns 9,791 AI Findings Into 55 Fixes
A GitHub–Microsoft pilot cut 9,791 raw AI security findings to 99 verified reports and shipped 55 fixes across six open source projects, with maintainers controlling every patch.

Why it matters
- GitHub reported on November 17, 2026 that the pilot produced 55 fixes across six open source projects as of November 15, 2026.
- AI triage and manual review cut 9,791 raw findings down to 99 private reports.
- Caddy's maintainers fixed all 22 reports they received in a week or less.
- Microsoft FORGE Lab researchers spent about two days per project checking findings by hand.
- GitHub will extend AI security scans and Copilot Max to alumni projects of the Secure Open Source Fund.
GitHub reported on November 17, 2026, that a security pilot run with Microsoft produced 55 fixes across six open source projects as of November 15, 2026. The trial, built on Microsoft's Codename MDASH system, filtered 9,791 possible issues down to 99 private reports, each checked by human researchers and accompanied by a proof of concept and a suggested fix.
Fourteen of those fixes had already landed in releases. The six projects were AutoGPT, Caddy, Express, FastMCP, ImageMagick and Netty. Maintainers decided which findings to fix, how to fix them and when to ship. Reports and patches stayed on GitHub throughout.
The numbers matter because cheap AI scanning has created a maintenance problem: duplicates and reports that are not actual vulnerabilities already consume maintainer time, while remediation still requires human judgment. GitHub wants researchers to consult maintainers before pointing another model at their projects and to submit patches alongside any findings.
How did the pipeline work?
Microsoft's Codename MDASH system scanned each project with its MAI Cyber model and suggested fixes. A GitHub Security Lab AI triage agent, built on tailored Copilot skills, then removed duplicates, checked validity and focused on high-impact bugs. Seven Microsoft FORGE Lab researchers manually checked every remaining finding before reports reached maintainers through GitHub private vulnerability reporting.
The design puts two constraints in sequence. Automation absorbs volume. Humans verify what automation cannot. GitHub described the project sample as small, the conclusions as early lessons and the manual review as not yet scalable — the FORGE researchers spent about two days per project checking findings by hand. Low-quality reports still reach maintainers every day.
What did maintainers actually do with the patches?
The first fixes shipped about a week after the initial scan, according to GitHub. Caddy's maintainers fixed all 22 reports they received in a week or less. ImageMagick, maintained by two people, also received 22 reports. In one case, its maintainers pushed a fix in 15 minutes after a summary identified the exact lines of code and showed that an apparently familiar issue was in fact different.
AutoGPT shows how the reports changed maintainer behavior. A report arrived three days after maintainers shipped a broken check, and it pointed to the right line. The maintainers took the fix further than the suggested patch, wrote it within a day and had it in production less than a week later. Across the pilot, maintainers used some suggested fixes as written and adapted others.
A Netty maintainer described the remaining constraint in plain terms: "Finding vulnerabilities is no longer our bottleneck. Fixing them is the hard part, and that depends on maintainers."
Why does this matter for AI-assisted disclosure?
The pilot's trade-off is concrete. Automation can reduce a flood of raw findings to a manageable private queue. Expert review and maintainer control determine whether those reports lead to patches. Remove either element and the system fails in a known direction: unreviewed AI reports return the triage burden to maintainers, while fixes without maintainer judgment break trust.
GitHub set out its recommended standard for AI-generated reports. Each finding should be checked against the project's existing knowledge, threat model and security policy. Reports should go privately. They should include a suggested fix. Maintainers should set the pace.
GitHub also identified two remaining needs. First, private-fork CI/CD testing would verify fixes on private forks before anything ships. Second, maintainer access to cyber-capable models would help maintainers verify a proof of concept before and after remediation — the same class of tool that found the bug, placed in the hands of the person fixing it.
What happens next?
GitHub plans to carry the pilot's lessons into private vulnerability reporting and security advisories. It will offer AI security scans with findings checked before private reporting, plus GitHub Copilot Max, to all alumni projects of the GitHub Secure Open Source Fund, and it will extend that support to future Fund sessions.
GitHub and Microsoft are also working to expand researcher review while preserving those checks. That is the open question the pilot leaves behind: the review layer that turned 9,791 findings into 99 credible reports took two researcher-days per project, and GitHub itself calls it not yet scalable. Whether expansion keeps that standard, or quietly drops it, will decide whether AI-assisted disclosure helps maintainers or simply resurfaces the flood.
Source: unite.ai
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
728 articles
Related articles
- GitHub and Microsoft's AI bug scanner produced 99 verified reports across six projects
- OpenAI's Daybreak Launches Patch the Planet for Open-Source Security
- Microsoft Patches 1,093 Flaws, Including Outlook Zero-Day Found With Claude
- Four Labs, One Blueprint: How AI Security Agents Converged in 2026
- Entire launches agent-first Git platform with automated evals