Finding flaws in open source has become cheap and fast; repairing them has not, and the repair work lands on maintainers. Many projects already receive AI-generated security reports in volume, including duplicates and items that were never vulnerabilities, so another raw scan adds burden rather than relief. Anyone aiming a model at someone else's project should clear it with the maintainers and arrive with fixes rather than findings. Christian Grobmeier, a Log4j maintainer, has framed the bar bluntly: the AI defending open source has to be better than the AI attacking it.
GitHub and Microsoft ran a joint pilot with maintainers to test whether AI could shift that balance — helping projects ship fixes instead of handing them more to triage. The vehicle was Codename MDASH, Microsoft's multi-model agentic scanning harness, applied to widely used open source software.
How a finding becomes a fix
The pilot had two ground rules: only findings that matter get sent, and no maintainer receives an unverified finding.
- Find.
Codename MDASH, an AI system from Microsoft Security, scanned each project with Microsoft'sMAI Cybermodel and proposed fixes for what it surfaced. - Filter. A GitHub Security Lab AI triage agent, built on tailored Copilot skills, removed duplicates, checked validity and concentrated on high-impact bugs, so that raw scan output never reached maintainers.
- Verify. Researchers from Microsoft FORGE Lab went through every surviving finding by hand.
- Report privately. Findings went out through GitHub private vulnerability reporting, each with a proof of concept and a suggested fix.
- Fix. Maintainers chose what to fix, how, and when.
Reports and fixes stayed on GitHub, the place maintainers already work.
Across the six projects, the system flagged thousands of candidate issues. After AI triage and manual review by seven FORGE Lab researchers, maintainers were sent 99 private reports. By November 15, 55 were fixed and 14 of those fixes had shipped in a release.


Early lessons from six projects
What follows reflects what the projects fixed and what their maintainers chose to share. Six projects is a small sample; treat these as early signals.
Fixing time is what gets saved
Maintainers spend their hours on fixes, so every report arrived pre-checked and bundled with a patch to review, adapt and merge. Work started from a scoped patch rather than a raw finding. The earliest fixes landed roughly a week after the first scan, and Caddy's maintainers cleared all 22 reports they received within a week or less, adopting the suggestions either as written or with edits.
For AutoGPT, a recent check of theirs had broken. The report arrived three days later and pointed at the right line; the maintainers extended the proposed patch beyond what was suggested, wrote their version in a day, and had it in production inside a week.
Finding vulnerabilities is no longer our bottleneck. Fixing them is the hard part, and that depends on maintainers.
Netty maintainer
Volume down, signal up
Maintainers received short lists instead of floods. Express maintainers said the reports got past the noise; Caddy maintainers said they exposed obscure bugs and closed gaps left by earlier reports; FastMCP's scan went further than any AI scanner had for that project. ImageMagick, run by just two people, took in 22 reports. The division of labor was consistent: AI found, researchers checked, maintainers remained in control.
A single ImageMagick report is illustrative. It resembled a class of issue the maintainers had seen before, but the summary showed it was different and named the exact lines. They pushed a fix in 15 minutes.
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.
ImageMagick maintainer
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.
Caddy maintainer
Unresolved gaps in the ecosystem
The pilot depended on careful manual work at each stage, and that work does not scale yet: seven researchers spent about two days per project on hand verification, while low-quality reports keep arriving at maintainers daily. Closing the distance to attackers requires a few things for open source: fixes delivered through private vulnerability reporting (PVR), CI/CD tests that maintainers can run on private forks to validate those fixes, and access to cyber-capable models so a proof of concept can be verified both before and after the fix.
The Codename MDASH scan provided unique insights into the security of our project above and beyond what AI scanners have provided us in the past
FastMCP maintainer
What the pilot asks of senders
For maintainers, the model changes the cost of a security report: a verified, private report with a suggested fix consumes less time, and triaging a scanner's output should never be the maintainer's job — it belongs to whoever sent it. For anyone using AI to hunt vulnerabilities in open source, the maintainer feedback here amounts to a standard worth copying: validate every finding first, against what the project already knows and against its threat model and security policy; report privately; include a suggested fix; and let maintainers set the pace.
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.
AutoGPT maintainer
Codename MDASH is very useful for cutting through the noise of vulnerability reports.
Express maintainer
Actions available now
- Maintainers should switch on private vulnerability reporting so researchers can report issues without exposing them publicly. Projects on GitHub can activate core security features in about two minutes with
gh-secure. - Organizations building on open source should enable Dependabot alerts to learn when a dependency needs a security update.
- Anyone sending AI-generated vulnerability reports should hold them to the standard described above before sending anything.
- Repositories with GitHub code security can lean on agentic autofix to remediate alerts automatically with Copilot.
Where this goes next
- Fixes in maintainers' existing workflow. GitHub is folding these lessons into private vulnerability reporting and security advisories, where GitHub Security Lab already works on cutting the burden of low-quality reports.
- MAI Cyber 2. Microsoft announced MAI Cyber 2; the pilot itself ran on the earlier MAI Cyber model. GitHub and Microsoft will continue the collaboration on open source security.
- Scaling review, starting with the GitHub Secure Open Source Fund. GitHub and Microsoft are working on scaling researcher review without compromising it. Every alumni project of the fund will be offered AI security scans whose findings are checked before being reported privately, along with GitHub Copilot Max, with the support extended to future sessions of the Fund.
The downstream effect is cumulative. Once a maintainer publishes a security advisory on GitHub and it is reviewed for the GitHub Advisory Database, Dependabot can warn every repository depending on that package. That logic underlies the GitHub Secure Open Source Fund: 216 projects so far, more than 300 maintainers, and more than 4,000 vulnerabilities fixed in software downloaded over 20 billion times a month.
Hear from maintainers about their experience with Codename MDASH.
Where the money comes from
Every project named in this pilot is an alumnus of the GitHub Secure Open Source Fund, which routes $10,000 to each project via GitHub Sponsors and pairs the funding with security training and an alumni community. The Fund, powered by GitHub Sponsors, targets open source that underpins the digital supply chain, drives innovation, and is critical to the AI stack. Companies that depend on such code can join as funding partners through the Fund page, where applications are also open.

Alongside the cash, the Fund supplies security training and Azure credits from Microsoft for Startups. The surrounding context is Microsoft's European Security Program and its Secure Future Initiative, whose stated principle is security above all else. The same logic extends downstream: if your software relies on open source, support the maintainers behind it.
The full alumni roster
The projects below span four broad areas of the stack.
AI, machine learning, and intelligent systems
ACI.dev, AlbumentationsX, ArviZ, AutoGPT/GravitasML, CAMEL-AI, Caracal, CocoIndex, Code-Graph-RAG, CodeCarbon, Cognee, Deep Agents, DIPY, DocsGPT, FastMCP, Jupyter, LadybugDB, LangChain, Matplotlib, n8n-MCP, Nasiko, NumPy, Ollama, ONNX, OpenBB Platform, OpenClaw, OpenCV, OpenMetadata, OpenSearch, PageIndex, pandas, Pydantic, PyMC, Ruby-OpenAI, Scenic, scikit-learn, SciPy, Serena, TokenOps, TraceRoot, Zeus.
Build systems, supply chain, and release tooling
Apache Airflow, Babel, Bluefin, bootc, Browserslist, Changesets, Colima, Concourse, Cucumber, the CycloneDX Python Library, Flux, Foundry, Gitoxide, golangci-lint, GoReleaser, hardhat-deploy, Jenkins, JReleaser, Jupyter Docker Stacks, NixOS/Nixpkgs, node-lru-cache, oapi-codegen, PostCSS, PyPI/Warehouse, rimraf, Task, Termux, Terra, Turborepo, Warpgate, webpack, zizmor.
Core languages, runtimes, and foundational libraries
Apache Commons, Apache Groovy, Byte Buddy, core-js, CPython, FS2, Gleam, Himmelblau, htmx, LLVM, modernc.org/sqlite, NAPI-RS, Node.js, Pkl, Pyodide, Rustls, termcolor, time-rs.
Developer tools and productivity platforms
AirQo, API Dash, ArduPilot, AssertJ, the AsyncAPI Initiative, Bevy, Bootstrap, calibre, Charset Normalizer, Cheerio, Ciphey, Cobra, CodeRunner, CSS Declaration Sorter, DataJourney, Diesel, DIGIT, Electron, ERPNext, fabric.js, GitPython, Hoppscotch, ImageMagick, jQuery, js-yaml, jsoup, JUnit, libyt, LORIS, MapStruct, Mastodon, Mathesar, Mautic, Mermaid, MessageFormat, Mockoon, NativeScript, Next.js, Nuxt, nvm, Oh My Zsh, p5.js, Path-to-RegExp, the Pelias Geocoder, Pest, Pillow, Polly, Proyecto Respira, PypeIt, python-benedict, qs, React Starter Kit, Readest, Resolve, Rocket.Chat, Selenium, shadcn/ui, Sphinx, Spyder, ssh_config, StackQL, Stirling PDF, Svelte, Thunderbird for Android, ToolJet, Two.js, Viper, virtualenv, Vuetify, Wagmi, WebdriverIO, xyflow, YAML, the Yii Framework, Yjs.
Web, networking, APIs, and infrastructure
actix-web, aiohttp, Apache APISIX, Apache Shiro, Apache Solr, Apache ZooKeeper, Arkime, Caddy, cmux, Cryptomator, curl, the CycloneDX .NET Library, CycloneDX cdxgen, Ente, etcd, evcc, Express, external-secrets, FastAPI, Fastify, Haraka, Helmet.js, Home Assistant, Hummingbird, jose, Keycloak, Keyshade, kgateway, Log4j, mimetype, NetBird, Netty, Oauth2 (Ruby), OAuthlib, openapi-stack, OpenWISP, PGPainless, quic-go, ScanAPI, ScanCode, Sniffnet, Social Auth App Django, Stalwart, Starlette, UAParser.js, urllib3, Vapor, varlock, Veramo, WebAuthn (Go), ZITADEL.



