FYI - ONNX will automatically exfiltrate data about your application with no way to opt out other than building it from source with telemetry disabled. Maybe you could block DNS too.
I don’t understand what you mean. I’ve maintained a decently popular ONNX wrapper for years, had 0 idea there was telemetry, and this is quite useful info to have, particularly because it would be a PITA to rebuild binaries right now.
I don't see the connection you're implying? What has ham radio got to do with any of this? There's nothing secret or being exfil'd in ham radio since it's all unencrypted and transmitted on public airwaves. In the USA anyway. I'm not sure what other licensing authorities in other countries allow.
> The ONNX Runtime (ORT) is the most widely used engine for serving deep
learning models in production. It is common for people to download ONNX
models from public repositories, such as Hugging Face, and to serve them to
their customers without too much scrutiny. In March 2026, we noticed a
parser flaw: an ONNX model can cause ORT to read from arbitrary file paths
(or memory addresses) without validations.
We found out about this this morning, when evals started failing in our pipelines. I've been looking around, but it doesn't seem like there was any advance warning about this.
Does anyone know why or how a repository with 22k stars just went dark?
That would be my guess also. About 6 months ago the azure-functions-host repo disappeared for a while (I got a concerning automated email from GitHub from my forked repo citing a ToS violation).
If I understand your implication, that's not what "supply chain attack" normally means. That term means attacking the mechanism that delivers your packages to deliver a different, more compromised, package. If the package itself is already compromised then that's just a straightforward dependency vulnerability.
Considering Microsoft co-developed ONNX with Facebook... I think that the more likely scenario is they took it dark after learning that it has security issues, but they should really have left a message or something, this is amateur hour tier. GitHub repos should let you leave a "Sorry we've hidden our repo, we learned about security issues" which kind of doesn't help anyone either... since now they have to find help elsewhere, and fall victim to sketchy 'forks' from other people...
I do wish I could pay some money for a package manager intermediary that "magically" handles things (for example defaulting to not promoting things until 2 weeks have passed), has better perf, and generally just does the right things.
Of course I wouldn't pay _that_ much money for it.
I don't get it, this is exactly how I have been mostly working since around 2010.
Nothing that lands on CI/CD pipelines is able to download anything directly from Internet, rather internal repos based on Nexus, Artfactory and similar.
Additionally everything that gets uploaded into them needs to have an approval process for the most part.
So, not only already exists, it is an intermediate process, and you can even pay depending on the features.
Such solutions exist. Some people are building it on their own. Some people just commit all their dependencies in their repository.
If you want it enforced for a whole team, with caching on top, that's
where registry proxies come in (Artifactory, Nexus, and similar). The big
ones take real effort to set up, while lighter ones may be all you need.
Vendoring your dependencies into the repo is the low-tech option that survives an upstream repo vanishing.
Disclosure: I work at a company which sells such solutions.
I’ve been using StableBuild for this sort of thing. It’s not exactly what you’re describing, but might be relevant.
You basically point your builds at it instead of the upstream repos. Once you’ve pulled something, it keeps that version around, even if it gets deleted or changed upstream.
Not sure if that covers what you need, but it’s worked well for me.
No that’s forking, vendoring usually means keeping a local copy of a library in /vendor that is automatically committed - rust and golang have first class support for this
The closest/furthest ahead fork I could find is https://github.com/intel/onnxruntime for anyone else needing it, though that's with Intel's OpenVINO modifications
It seemed to have stalled in investment before i think. Performance especially for CPU only models ( there are some use cases for this ) seemed to have lagged a lot, behind what properly hand optimized code could do.
The gpu work was mostly handed of e.g to TensorRT and that also seems since the push for more general LLM models have not really improved.
Aren't there other ONNX runtimes out there? A quick google search shows a few written in Rust: Tract, RTen, WONNX, Burn&Candle. Also, IIRC there's a Julia-based one as well.
I maintain RTen (https://github.com/robertknight/rten). It is a good drop-in replacement for ONNX Runtime for many use cases where you are running inference on CPU (there is no GPU support yet). There is no Python or NodeJS API at present, though you could get an agent to build a small wrapper quite easily.
I've used your library ocr to read serial numbers on an IoT device on a mobile phone. It was really responsive and faster than tesseractjs in a web worker. Just wanted to say thanks for such well thought software!
Have used it with wasm and it's lovely, but I was kindof disappointed to learn internally it's basically another serialisation format. I imagined gigabytes of generated Rust code hehe
The graph is converted into Rust code that calls Burn APIs but the weights are converted into Burn's format, called burnpack (similar to safetensors but instead of JSON we use CBOR for weights meta data).
Not to take away from the fact that I don't like that this was done, but in my experience development on ONNX has been pretty slow recently. I moved away from a few months back because I felt bug fixes were coming in pretty slow, despite being reported.
How odd. https://onnxruntime.ai is still up, and the packages still appear to be on PyPI, Maven Central etc. Guess I won't submit the PR I was working on tonight then.
Microsoft Agent Framework has been compromised[0]. The repo is also offline[1]. Possibly related?
[0] https://socket.dev/blog/agent-framework-pypi-package-comprom...
[1] https://github.com/microsoft/agent-framework
Without any official communications, Occam's razor means we can assume at least one admin account in Microsoft GitHub org got compromised.
That assumption means Occam’s Razor doesn’t apply.
If you were looking for fewer assumptions then the conclusion would just be that an exploit got missed during the PR step.
Or a contributor to both of these related repos?
Why are you assuming a compromise? Microsoft Ignite is happening and they announced a bunch of stuff. They're probably consolidating repos.
FYI - ONNX will automatically exfiltrate data about your application with no way to opt out other than building it from source with telemetry disabled. Maybe you could block DNS too.
onnxruntime.disable_telemetry_events() works doesn't it?
Same with setting `ORT_DISABLE_TELEMETRY=1` in the environment before loading the library.
Not to be that guy but I don't think this is the gotcha y'all think it is. I didn't know that ONNX had telemetry.
edit: I seem to have misunderstood the intent of the parent comments. Apologies
Top-level comment said "no way to opt out other than building it from source with telemetry disabled", so if that's not true then this is important to know. See https://www.astralcodexten.com/p/if-its-worth-your-time-to-l...
I'm just passing information for anyone who wants to disable it without recompiling anything.
I don’t understand what you mean. I’ve maintained a decently popular ONNX wrapper for years, had 0 idea there was telemetry, and this is quite useful info to have, particularly because it would be a PITA to rebuild binaries right now.
Isn't ONNX itself not MS specific?
They probably mean "ONNX Runtime", which has unfortunately become synonymous with ONNX-the-standard.
Yikes. Another bad actor who thinks they are entitled to data.
Much like ham radio.
I don't see the connection you're implying? What has ham radio got to do with any of this? There's nothing secret or being exfil'd in ham radio since it's all unencrypted and transmitted on public airwaves. In the USA anyway. I'm not sure what other licensing authorities in other countries allow.
Maybe related to one of their packages that was compromised? Not sure if GitHub was the attack vector, but possibly tidying stuff up
Source: https://opensourcemalware.com/blog/agent-framework-pypi-comp...
I wonder if they found new vectors that backdoor'ed models could exploit, like that discussed in the most recent Phrack [^0]
[0]: https://phrack.org/issues/73/backdoor-attacks-against-machin...
> The ONNX Runtime (ORT) is the most widely used engine for serving deep learning models in production. It is common for people to download ONNX models from public repositories, such as Hugging Face, and to serve them to their customers without too much scrutiny. In March 2026, we noticed a parser flaw: an ONNX model can cause ORT to read from arbitrary file paths (or memory addresses) without validations.
We found out about this this morning, when evals started failing in our pipelines. I've been looking around, but it doesn't seem like there was any advance warning about this.
Does anyone know why or how a repository with 22k stars just went dark?
Probably being used in supply chains, and got attacked.
That would be my guess also. About 6 months ago the azure-functions-host repo disappeared for a while (I got a concerning automated email from GitHub from my forked repo citing a ToS violation).
If I understand your implication, that's not what "supply chain attack" normally means. That term means attacking the mechanism that delivers your packages to deliver a different, more compromised, package. If the package itself is already compromised then that's just a straightforward dependency vulnerability.
Why do you think star amount has anything to do with ability to go dark?
From a pure lumens perspective it seems obvious.
it's called "Open Neural Network Exchange (ONNX)", perhaps microsoft is taking a page from OpenAI on what open meanings.
Considering Microsoft co-developed ONNX with Facebook... I think that the more likely scenario is they took it dark after learning that it has security issues, but they should really have left a message or something, this is amateur hour tier. GitHub repos should let you leave a "Sorry we've hidden our repo, we learned about security issues" which kind of doesn't help anyone either... since now they have to find help elsewhere, and fall victim to sketchy 'forks' from other people...
This is why we vendor everything.
You never know when it's going to vanish
I do wish I could pay some money for a package manager intermediary that "magically" handles things (for example defaulting to not promoting things until 2 weeks have passed), has better perf, and generally just does the right things.
Of course I wouldn't pay _that_ much money for it.
I don't get it, this is exactly how I have been mostly working since around 2010.
Nothing that lands on CI/CD pipelines is able to download anything directly from Internet, rather internal repos based on Nexus, Artfactory and similar.
Additionally everything that gets uploaded into them needs to have an approval process for the most part.
So, not only already exists, it is an intermediate process, and you can even pay depending on the features.
Doesn't Artifactory provide exactly that? Along with having a bunch of features that are generally the right things.
Though I'm pretty sure their enterprise pricing is "give us a call." Maybe your last sentence was a nod to this.
Such solutions exist. Some people are building it on their own. Some people just commit all their dependencies in their repository.
If you want it enforced for a whole team, with caching on top, that's where registry proxies come in (Artifactory, Nexus, and similar). The big ones take real effort to set up, while lighter ones may be all you need.
Vendoring your dependencies into the repo is the low-tech option that survives an upstream repo vanishing.
Disclosure: I work at a company which sells such solutions.
I’ve been using StableBuild for this sort of thing. It’s not exactly what you’re describing, but might be relevant.
You basically point your builds at it instead of the upstream repos. Once you’ve pulled something, it keeps that version around, even if it gets deleted or changed upstream.
Not sure if that covers what you need, but it’s worked well for me.
I have not tried it myself: https://github.com/artifact-keeper
If it were me, I would take the hint and maybe not trust a vendored version right now?
I believe vendoring in this context means forking it and serving it from your own repo.
No that’s forking, vendoring usually means keeping a local copy of a library in /vendor that is automatically committed - rust and golang have first class support for this
But if we think they took it down because of security issues…
The closest/furthest ahead fork I could find is https://github.com/intel/onnxruntime for anyone else needing it, though that's with Intel's OpenVINO modifications
Sourceforge got your back https://sourceforge.net/projects/onnx-runtime.mirror/files/
Note to people downloading these archives, there is a high likelihood that they contain malware.
It seemed to have stalled in investment before i think. Performance especially for CPU only models ( there are some use cases for this ) seemed to have lagged a lot, behind what properly hand optimized code could do.
The gpu work was mostly handed of e.g to TensorRT and that also seems since the push for more general LLM models have not really improved.
As a former (tiny patch, pre-AI) contributor, didn't they have a license that requires everybody's approval to do this?
No. Even if a license required all contributors to agree to change the license, there's no requirement to keep it available on Github.
Thanks!
Aren't there other ONNX runtimes out there? A quick google search shows a few written in Rust: Tract, RTen, WONNX, Burn&Candle. Also, IIRC there's a Julia-based one as well.
I maintain RTen (https://github.com/robertknight/rten). It is a good drop-in replacement for ONNX Runtime for many use cases where you are running inference on CPU (there is no GPU support yet). There is no Python or NodeJS API at present, though you could get an agent to build a small wrapper quite easily.
I've used your library ocr to read serial numbers on an IoT device on a mobile phone. It was really responsive and faster than tesseractjs in a web worker. Just wanted to say thanks for such well thought software!
https://github.com/tracel-ai/burn-onnx
I maintain this project at the moment.
It convert ONNX models into native, backend-agnostic Burn code for inference and fine-tuning.
This looks awesome, i may be using this in prod very soon. Ive had some ONNX issues this looks to solve
Have used it with wasm and it's lovely, but I was kindof disappointed to learn internally it's basically another serialisation format. I imagined gigabytes of generated Rust code hehe
The graph is converted into Rust code that calls Burn APIs but the weights are converted into Burn's format, called burnpack (similar to safetensors but instead of JSON we use CBOR for weights meta data).
Not to take away from the fact that I don't like that this was done, but in my experience development on ONNX has been pretty slow recently. I moved away from a few months back because I felt bug fixes were coming in pretty slow, despite being reported.
What do you use instead?
https://github.com/syntax-syndicate/onnxruntime
For others:
> https://github.com/syntax-syndicate/onnxruntime
Check the web archive link for commit hashes, the syntax-syndicate fork seems to be the remaining legit fork.
[1] https://web.archive.org/web/20261113082932/https://github.co...
Maybe related to the recent release of MAI? Or the Microsoft AI models?
agent-framework is down as well:
https://github.com/microsoft/agent-framework/
How odd. https://onnxruntime.ai is still up, and the packages still appear to be on PyPI, Maven Central etc. Guess I won't submit the PR I was working on tonight then.
Microsoft, please, I can’t have popcorn every day.
omg, I just built something depending on this. Hopefully the web version is still cached in NPM.
Someone with a copy should soft fork it
Better delete all the open source matrix multipliers because we're making it illegal
Very concerning!
The Maven package is still available, and the most recent version is from a few days ago:
https://mvnrepository.com/artifact/com.microsoft.onnxruntime...
Now Jensen's gone they can start dismantling the infrastructure that allows people to run models on their own hardware.
So... all of these also vanished.
(possible they were gone earlier, but e.g. onnxruntime-extensions had a push yesterday)It's here: https://github.com/onnx/onnx
The onnx repo is just a spec. Microsoft's ONNX Runtime is an implementation of the spec.
This is not the same package.