This closes remaining cross-origin video paths in WebGPU’s importExternalTexture that were not covered by the origin-taint check, defeating the Same-Origin Policy for video pixels. GPUDevice::importExternalTexture ingests an HTMLVideoElement’s frames as a WebGPU external texture that a shader can sample and read back. To respect the SOP (as canvas does by tainting), an origin-taint check must reject a cross-origin, non-CORS video before it enters the pipeline.

Pre-patch, importExternalTexture had paths that imported the video WITHOUT that check — notably the Cocoa fast path that reuses/undestroys a previously-created external texture (externalTextureForDescriptor) and the main import path.

The fix adds a checkVideoElementOriginTaint lambda that fetches the caller’s security origin from scriptExecutionContext()->securityOrigin() and, if videoElement.taintsOrigin(securityOrigin), returns a SecurityError (‘Cross origin external videos are not allowed in WebGPU’); it invokes this check on BOTH the reused-external-texture path and the main path before importing.

The restored invariant is that a cross-origin tainted video cannot enter (or be reused in) a WebGPU external texture through any importExternalTexture path, so its pixels cannot be sampled and exfiltrated. This is an information-disclosure fix, not memory corruption.

A cross-origin information disclosure in the WebContent process defeating the Same-Origin Policy for video: uncovered importExternalTexture paths let a page read pixels of a cross-origin video. No memory corruption; the risk is exfiltration of cross-origin visual content (e.g. authenticated video), reliable and stealthy.