ONNX Runtime vanishes from GitHub: what breaks, what works, and why mirrors are risky
Since the evening of November 17, GitHub has served nothing but error pages for ONNX Runtime, Microsoft's AI runtime library, and that includes every release download. If you install via pip, NuGet or conda, you won't notice a thing. If you pull binaries straight from GitHub, you have a problem, and many projects are now turning to substitute sources whose origin is hard to verify.

What's going on
Since the evening of November 17, ONNX Runtime has been unreachable on GitHub. github.com/microsoft/onnxruntime returns nothing but a 404, and that applies to the code, the releases page, the API and every single download file. ONNX Runtime is a runtime library: the piece of software that actually executes a trained AI model in ONNX format, on the CPU, the GPU or an AI chip (NPU). It sits inside text-to-speech engines, speech recognition, background removers, in-browser AI and Windows itself.
Most coverage so far asks what happened to the repository. We asked a different question: How much local AI depends on this one library, which distribution channels still work, and what risks arise when the community starts building its own substitute sources?
What we know and what we don't
A Microsoft spokesperson told The Register:
"We have made some repositories private while we investigate malicious content."
He didn't say which repositories. According to Hacker News and RuntimeWire, at least seven others have disappeared alongside the main repo, including onnxruntime-genai, onnxruntime-extensions, the Swift package, the GitHub Actions templates and sample collections. The ONNX specification at onnx/onnx is still online; it's a separate project. One maintainer wrote in issue #8842 there that they know "no more than the public."
The timing is striking:
| Time (UTC) | Event |
|---|---|
| Nov 16, 23:39 | last successful CI run with a GitHub download (trusty-tools) |
| Nov 17, 20:35 | poisoned agent-framework 1.26.1 appears on PyPI |
| Nov 17, 21:06 | PyPI quarantines 1.26.1 |
| Nov 17, 21:08 | 1.26.2 follows 2 minutes 23 seconds later (StepSecurity) |
| Nov 17, approx. 22:20 | onnxruntime repo returns 404 (per nuzky and byteiota) |
| Nov 18, from approx. 04:15 | first reported installer failures (FrameEver, trusty-tools) |
Agent-framework is Microsoft's Python package for building AI agents. According to OX Security, SafeDep, Wiz and JFrog, both tampered versions contained a credential stealer. A .pth file, one of Python's startup mechanisms, ran it every time Python started, even if the package was never imported. It harvested environment variables, key files and tokens for Azure, Azure DevOps and GitHub Actions, and sent them to a webhook.site address.
How the attacker got in is disputed. The Register reports, with caveats, a hijacked admin account in Microsoft's GitHub organization. SafeDep found a second PyPI maintainer named vibiret, created on the day of the attack, though it's unconfirmed whether that account uploaded the packages. Wiz believes a client certificate for ESRP, Microsoft's internal release pipeline, was compromised. JFrog found that the wheel was packaged by a local script and didn't come from Microsoft's pipeline. Important: Neither Microsoft nor the security researchers have so far linked the ONNX Runtime outage to the attack. Nor has anyone reported that ONNX Runtime itself was tampered with.
This isn't the first time. In June 2026, according to The Hacker News and TechCrunch, 73 Microsoft repositories were infected with the "Miasma" worm and taken offline for days. Some came back only "after review"; others stayed offline longer.
Which channels still work
We queried the official package sources ourselves on November 18:
| Channel | Status | Notes |
|---|---|---|
| GitHub repo and releases | 404 | old versions too (1.20.1 through 1.31.1 reported) |
PyPI onnxruntime / onnxruntime-gpu |
works, 1.32.0 from Nov 13 | no provenance attestations, no Trusted Publishing |
NuGet Microsoft.ML.OnnxRuntime |
works, 1.32.0 | signed package, 19.1M downloads, 427 dependent packages |
npm onnxruntime-web / -node |
works, 1.31.1 | registry signatures, no provenance attestation |
| conda-forge | works, 1.32.0 from Nov 14 | 3.5M downloads |
| download.onnxruntime.ai | partially | homepage only shows old training packages; per Spoken-Margins, the Swift binaries are still being served |
| Docs at onnxruntime.ai | reachable | — |
The numbers show how much depends on this library. According to pypistats, onnxruntime was downloaded around 141 million times last month, plus another 2.7 million for onnxruntime-gpu. On npm, in the week of November 11–17, onnxruntime-node hit 9.7 million downloads and onnxruntime-web 8.1 million. Most of that is likely automated CI runs, not humans. That doesn't make it any less serious: these are exactly the pipelines that are now breaking or fetching their files from new sources.
Who breaks and who keeps running
Our take after going through the projects: what matters isn't whether a program uses ONNX Runtime, but where it gets the library from.
Keeps running: Anything installed through a package manager. The text-to-speech engine kokoro-onnx pulls onnxruntime via pip; the background remover rembg uses the [cpu] and [gpu] extras. The speech recognizer sherpa-onnx downloads its own builds from the csukuangfj/onnxruntime-libs repo and is therefore unaffected. According to its docs, transformers.js loads the browser WASM files from jsDelivr by default, which is fed from npm. Windows ML doesn't depend on the GitHub repo either. Microsoft describes it as "the Windows-supported and maintained copy of ONNX Runtime," distributed as a system-wide component via the Windows App SDK. Programs that are already installed keep working regardless. Only fresh installs fail.
Breaks: Anything that downloads release archives directly from GitHub. Reported cases include:
- YuE2-Studio and ACE-Step-Studio (AI music): the synced-lyrics package won't load, with the error "ONNX Runtime 1.31.1 … status 404".
- Speech Studio (speech AI): first-time setup on Windows fails on ORT 1.28.2.
- nuzky (face and subject detection): per PR #418, all builds, checks and installer releases were blocked.
- mdoc (OCR), FrameEver, openscreen, gortex and AIRI: CI pipelines and setup scripts are down.
- Node.js projects on Linux:
onnxruntime-nodeuses a postinstall script, a program npm runs automatically after installation, to fetch CUDA files from GitHub. In the FailSafe project, this brokenpm cithrough the chain piper-tts-web → @huggingface/transformers → onnxruntime-node. The workaround is the environment variableONNXRUNTIME_NODE_INSTALL_CUDA=skip. - Swift projects: Spoken-Margins could no longer reach the deleted Swift package repo, leaving 68 pull requests blocked.
Substitute sources are the next risk
The community is helping itself quickly, but in very different ways. And that's exactly where the problem lies.
Option 1: NuGet as a parts warehouse. Many projects now extract the native libraries from Microsoft's NuGet package and pin the package with a fixed SHA-256 checksum, a kind of fingerprint for the file. This is the cleanest solution, because the file is signed by Microsoft. But the mdoc developer found a catch. The Windows DLL in the NuGet package is byte-identical to the GitHub version; the macOS library is not. It carries the same Apple developer ID, but it's a separate build with a different internal identifier, and 304,583 bytes differ. If you pinned old checksums, you'll have to replace them. Moments like this are when checks tend to get switched off "temporarily."
Option 2: Self-hosted mirrors. A mirror is a copy of files hosted somewhere else. Speech Studio hosts four Windows archives with SHA-256 values under the release tag onnxruntime-mirror. Three of them are said to be byte-identical to Microsoft's originals. The fourth, the CUDA 13 variant of 1.28.2, was repackaged from the 41 files of an existing installation. According to the maintainer, every DLL carries a valid Microsoft signature. The fix is a PowerShell one-liner that runs a script straight from the web. It's well-intentioned and transparent. But the checksums come from the mirror operator, and the original to compare against is offline.
Option 3: Copies from the cache. Spoken-Margins republished the Swift repo from a Mac's local package cache. A bot flagged the merge because the usual checks were skipped. The only evidence that the copy matches the original is the pinned commit hash.
Option 4: Third-party sources. On Hacker News, a SourceForge mirror, Gitee mirrors from China, forks and GitHub proxy sites are circulating. One user warned that such archives could contain malware. There's no evidence of that so far, but there's also no easy way to rule it out.
Our take: an attacker wouldn't even need access to Microsoft. Offering a credible substitute source in an emergency is enough. In the Kokoro case we covered, an expired domain was all it took to hijack a model with 13 million downloads. Here, a GitHub release with the right file name would do. Add to that a finding we checked on PyPI ourselves: the ORT wheels for 1.32.0 have no provenance attestations, and the upload client is listed as RestSharp/106.13.0.0. StepSecurity found the same client on the poisoned agent-framework uploads. That proves nothing about ONNX Runtime. But it shows that you can't independently verify where these packages came from; you have to trust the internal pipeline.
What developers and users should do now
- Don't download binaries directly from GitHub. Use PyPI, NuGet, npm or conda-forge instead, and in a company setting, go through a proxy cache like Artifactory or Nexus.
- Pin versions and checksums: with
pip install --require-hashes, lockfiles for npm, yarn or uv (files that record exact versions and hashes), and SHA-256 pins for archives. Checksums from a mirror only help if they come from an independent source. - Verify signatures, not just hashes. NuGet can enforce Microsoft-signed packages with
signatureValidationMode=requireandtrustedSigners. For DLLs, the Authenticode signature, a digital signature from the vendor embedded in the file, shows whether "Microsoft Corporation" signed it. On macOS, the same goes for the Developer ID. - Document workarounds and roll them back later, such as
ONNXRUNTIME_NODE_INSTALL_CUDA=skip, instead of disabling checks across the board. - If you're an end user: installed programs keep working. Don't fix anything with one-liners from forums. Wait for an update from the developer.
- If you installed agent-framework 1.26.1 or 1.26.2, every analysis agrees: treat all reachable credentials as stolen and rotate them. Version 1.26.0 is safe.
The bigger picture
The outage shows that GitHub Releases have quietly become infrastructure for many projects, even though GitHub was never designed for that. As with left-pad in 2016, a single deleted source makes thousands of builds fail. But as with xz-utils in 2024, the real danger lies less in the outage than in trusting the substitutes. Microsoft hasn't said whether, when or in what state the repo will return. It's also unclear what the "malicious content" was. Until Microsoft explains, you shouldn't use even a restored repo without verifying it.
Our take: the damage to end users has been small so far, because the package channels are up. The bigger problem is the precedent. One of the most widely used AI libraries has had two incidents in six months and still offers no verifiable provenance on PyPI. Projects that switch to signed NuGet packages and pinned hashes now will come out more robust than before. Those that simply plug in the next download URL they find have just traded an outage for a supply chain risk.
Sources
- The Register: Microsoft's official Python agent framework hijacked theregister.com
- SecurityWeek: Microsoft Agent Framework Releases Poisoned securityweek.com
- SafeDep: agent-framework Compromised Twice on PyPI safedep.io
- OX Security: Hijacked releases of agent-framework ox.security
- Wiz: Malicious version of agent-framework published to PyPI wiz.io
- StepSecurity: agent-framework PyPI Package Compromised stepsecurity.io
- JFrog Security Research: agent-framework hijack research.jfrog.com
- byteiota: Microsoft's ONNX Runtime Is Gone from GitHub byteiota.com
- GitHub: onnx/onnx Issue #8842 github.com
- GitHub: timoncool/YuE2-Studio Issue #271 github.com
- GitHub: timoncool/speech-studio Issue #41 github.com
- GitHub: speech-studio Release onnxruntime-mirror github.com
- GitHub: lawkitt/mdoc PR #939 (NuGet-Fallback, Hash-Vergleich) github.com
- GitHub: nuzkyapp/nuzky PR #418 github.com
- GitHub: bobmatnyc/trusty-tools Issue #11458 github.com
- GitHub: AniBullet/FrameEver Issue #178 github.com
- GitHub: MythologIQ-Labs-LLC/FailSafe Issue #704 github.com
- GitHub: berkmancenter/llm_engine Issue #457 github.com
- GitHub: Spoken-Margins PR #4568 github.com
- GitHub: getopenscreen/openscreen Issue #1625 github.com
- GitHub: zzet/gortex Issue #1491 github.com
- PyPI: onnxruntime 1.32.0 (Dateien) pypi.org
- PyPI: onnxruntime-gpu pypi.org
- pypistats: onnxruntime pypistats.org
- NuGet: Microsoft.ML.OnnxRuntime nuget.org
- npm-Registry: onnxruntime-web latest registry.npmjs.org
- conda-forge: onnxruntime anaconda.org
- Microsoft Learn: What is Windows ML? learn.microsoft.com
- sherpa-onnx: CMake-Download für onnxruntime github.com
- The Hacker News: Microsoft restores some GitHub repos (Juni 2026) thehackernews.com
- pip-Doku: Secure installs (--require-hashes) pip.pypa.io
- Microsoft Learn: NuGet package trust boundaries learn.microsoft.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.