GitHub and Microsoft test AI that sends fix PRs, not just bug alerts

- Reading
- 5 min
In a pilot with six open source projects, GitHub and Microsoft used an AI scanner verified by human researchers to deliver private reports with a suggested fix. Maintainers accepted 55 of the 99 fixes they received.
The problem isn't finding bugs, it's having time to fix them
Since AI tools made it cheaper to scan code for vulnerabilities, maintainers of open source projects have been hit with a flood of automated reports, many of them duplicates or simply wrong. Finding potential flaws became fast and cheap. Fixing them still requires human judgment, and that work falls on whoever maintains the project, usually without pay and without spare time.
The pressure is real. Christian Grobmeier, maintainer of Log4j, summed up the problem this way:
The AI defending open source has to be better than the AI attacking it.
Christian Grobmeier, maintainer of Log4j
It was out of this frustration that GitHub and Microsoft, according to a post published on the GitHub Blog on November 17, 2026, decided to test a different question: instead of sending more findings, can AI be used to send ready-made fixes?
How the pilot worked
The tool behind the experiment is Codename MDASH, described by Microsoft as a multi-model agentic scanning "harness." The pipeline had five steps, designed so that no raw finding would reach a maintainer without being checked:
- Find: Codename MDASH scanned each project using the MAI Cyber model, from Microsoft Security, and already suggested fixes for what it found.
- Filter: a triage agent from GitHub Security Lab, built on Copilot skills, removed duplicates and discarded invalid findings, isolating the highest-impact bugs.
- Verify: researchers from Microsoft FORGE Lab manually checked each remaining finding.
- Report privately: the reports reached maintainers through GitHub's private vulnerability reporting feature, already including a proof of concept and a suggested fix.
- Fix: the decision of what, how, and when to fix remained entirely with the maintainers.
The entire process, from report to merge, happened inside GitHub, where maintainers already work.
The numbers: from thousands of findings to 55 accepted fixes
Codename MDASH flagged thousands of potential issues across the six tested projects. After AI triage and manual review by seven Microsoft FORGE Lab researchers, that volume dropped to 99 private reports sent to maintainers.
By November 15, 2026, maintainers had already fixed 55 of those 99 reports, more than half, and 14 of the fixes were already in a published release. The first fixes shipped about a week after the initial scan.
In short: the bottleneck hasn't moved (it's still the maintainer's time), but the cost of each report has dropped, because it arrives already checked and with a proposed solution, not as a raw data dump to triage.
What happened in each project
The six pilot projects, all former grantees of the GitHub Secure Open Source Fund, had distinct experiences:
- Caddy: received 22 reports and fixed all of them within a week or less, using the suggested fixes, some as-is and others with adjustments.
- AutoGPT: had just shipped a check that was broken. The report arrived three days later, pointing to the exact line. The team went beyond the suggested patch, wrote the fix in a day, and put it into production in under a week.
- ImageMagick: maintained by a team of just two people, received 22 reports. One of them looked like a type of bug already seen before, but the AI-generated summary showed it wasn't, pointing to the exact lines. The fix shipped in 15 minutes.
- FastMCP: according to the maintainer, the scan went beyond what previous AI scanners had already delivered to the project.
- Express: maintainers said the reports cut through the noise of irrelevant findings.
- Netty: summed up the gain in the sentence that opens this piece, about the bottleneck no longer being finding vulnerabilities.
The ImageMagick maintainer compared the time before and after:
Normally, a security advisory would take us around 30 minutes to 1 hour on average to read, understand, and make a fix. It is a huge difference.
Maintainer of ImageMagick
The Caddy maintainer, meanwhile, highlighted the quality of the findings themselves, not just the speed:
Codename MDASH was able to find unknown, obscure security bugs in our project with succinct reports, and it filled in gaps from some prior reports that hadn't been noticed.
Maintainer of Caddy
What still doesn't scale
The GitHub Blog post itself admits the pilot's limit: seven researchers spent about two days per project checking findings by hand. That human verification work is what guarantees the reliability of the reports, but it doesn't scale as it stands today, while low-quality reports keep arriving at maintainers' doors every day, from other sources.
According to the post, for this model to work at scale, open source needs three things: delivering fixes via private vulnerability reporting (PVR), the ability to run CI/CD tests on private forks to validate fixes, and maintainer access to models with cyber capability to verify proofs of concept before and after the fix.
Why this matters for anyone who maintains (or depends on) open source code in Brazil
For anyone maintaining an open source project, the practical shift is this: a checked, private report with a suggested fix takes far less time than a raw scanner dump. The bar that GitHub Security Lab itself proposes, based on feedback from the six maintainers, is clear: anyone sending an AI-generated vulnerability report should check each finding against the project's threat model and security policy, report it privately, and include a suggested fix, leaving the pace in the maintainer's hands.
For anyone building on top of open source dependencies, the post's practical advice is to turn on Dependabot to know when a dependency in use needs a security update. For anyone maintaining a repository on GitHub, basic security features can be turned on in about two minutes with gh-secure, and private vulnerability reporting can be enabled so researchers can report without making the flaw public.
The move also ties into an announcement made the same day: Microsoft launched MAI Cyber 2, the successor to the model that ran the entire Codename MDASH pilot. GitHub and Microsoft say they will carry this learning over to private vulnerability reporting and to GitHub's security advisories, and that they will offer AI security scans, already checked by researchers, to all former grantees of the GitHub Secure Open Source Fund, which today totals 216 projects and more than 300 maintainers fixing vulnerabilities in software downloaded more than 20 billion times a month.
The AutoGPT maintainer summed up the takeaway for anyone building and maintaining software today:
Features are easier and easier to ship, but in the future maintaining software will probably come down to who cares enough to manage the security.
Maintainer of AutoGPT
Translated from the Brazilian Portuguese original · Read the original
The EU's CRA turns supply chain security into a code obligation for anyone exporting to the bloc
An InfoQ podcast with OpenSSF's CTO details how tools like Scorecard and SLSA became a practical path to complying with the Cyber Resilience Act, the European law that already affects anyone exporting software to the bloc.





