GitHub and Microsoft pilot Codename MDASH to automate security vulnerability fixes
Microsoft and GitHub are testing an AI-driven pipeline that delivers verified security patches rather than raw vulnerability reports to open-source maintainers.
Summary
GitHub and Microsoft have concluded a pilot program using a system called Codename MDASH to address the growing burden of AI-generated bug reports on open-source maintainers. Instead of flooding repositories with unverified findings, the initiative focuses on delivering high-quality, actionable security patches through a multi-stage verification process.
The pilot involved six major open-source projects, including Caddy, Express, and ImageMagick. By utilizing Microsoft's MAI Cyber model and human researchers from the FORGE Lab, the system successfully filtered thousands of potential issues down to 99 high-impact, private reports, many of which were resolved by maintainers within a week.
The Codename MDASH workflow
- Multi-stage Pipeline: The process utilizes Microsoft's MAI Cyber model for initial scanning, followed by an AI triage agent from GitHub Security Lab to remove duplicates and validate findings.
- Human-in-the-loop Verification: Every finding identified by the AI is manually verified by researchers from the Microsoft FORGE Lab before being presented to project maintainers.
- Private Reporting Mechanism: Validated vulnerabilities and suggested fixes are delivered via GitHub private vulnerability reporting to prevent zero-day exploitation during the patching process.
- Actionable Output: Rather than providing raw scan data, the system generates a proof of concept and a suggested code patch, allowing maintainers to review, adapt, and merge fixes directly into their workflows.
Why it matters
This approach shifts the AI paradigm from increasing the volume of noise to increasing the velocity of security patches. For the open-source ecosystem, this could mitigate maintainer burnout caused by automated scanning tools and improve the overall security posture of critical infrastructure like Netty and ImageMagick.
Original reference: github.blog →