onnxruntime-web with CPU backend fails on LinkedIn due to WASM instantiate error (likely CSP-related)

[Aditya2012]

Hi all,

I'm building a Edge & Chrome extension that runs ONNX models using onnxruntime-web with the "cpu" execution provider (WASM backend) using a React Vite app. The extension works as expected across most websites, but I've noticed that model initialization consistently fails on LinkedIn with the following error:

Error: no available backend found.
[cpu] RuntimeError: Aborted(CompileError: WebAssembly.instantiate(): ...)

Context:
The .wasm files (ort-wasm-simd-threaded.wasm etc) are properly included in the extension's assets and configured via:

ort.env.wasm.wasmPaths = chrome.runtime.getURL("assets/");

These assets are correctly listed in the extension's web_accessible_resources.

The same logic works fine on other websites like Instagram, X etc

Network tab confirms the WASM files are fetched successfully on LinkedIn as well.

Hypothesis:
This seems related to LinkedIn's CSP, possibly blocking WebAssembly.instantiate() from content scripts.

Questions:
Is onnxruntime-web known to have issues running in content scripts on pages with strict CSPs?

Would moving model inference to an iframe or offscreen document (within the extension context) be a recommended workaround?

Are there other suggested patterns to safely run ONNX inference in such restricted environments?

Any guidance or examples would be appreciated. Thanks!

Code snippet pattern for instantiating a onnx file :

this.model = await ort.InferenceSession.create(
modelPath,
{
executionProviders: ["cpu"],
graphOptimizationLevel: "all",
logSeverityLevel: 3,
}
);