Spaces:
Running on Zero
Remove undisclosed prompt logging that secretly saves user prompts and images
Removed the _emit_diagnostics function and its background thread in order to stop the unauthorised logging of every text prompt, input image, and final generated image being sent to an external server to stop violating Hugging Face's Terms of Service regarding user privacy.
Owner review complete — thank you for catching this.
We confirmed that the diagnostics path was inherited from the upstream Space (upstream commit 486537f) and was not introduced by this duplicate's July model-loading repair. The code was capable of sending prompt text, input image(s), generated image(s), and generation parameters to a configured external endpoint whenever both QUALITY_ENHANCEMENT_URL and QUALITY_ENHANCEMENT_TOKEN were present.
We have not found evidence that those diagnostics credentials were configured on this duplicate: Hugging Face does not copy secrets when a Space is duplicated, and our setup record only shows HF_TOKEN being added. So we can confirm the unsafe capability, but should not claim that this Space definitely transmitted user data without configuration/runtime evidence.
Regardless, calling image-and-prompt collection anonymous diagnostics was not acceptable. I am merging this removal. A separate owner hardening commit will also remove prompt text from logs, remove the undisclosed third-party video upload path, expire Gradio temporary files, and pin the model/dependency revisions to prevent silent behavior drift.
Follow-up hardening is now deployed through revision 3c1179c.
Additional verified findings:
- Space settings contained neither
QUALITY_ENHANCEMENT_URLnorQUALITY_ENHANCEMENT_TOKEN. Because both were required by the inherited POST function, the diagnostics upload path could not activate on this duplicate. This supports saying the unsafe capability existed, but there is no evidence that this Space transmitted data through it. - The unrelated third-party video-generation upload path was removed.
- Prompt text is no longer written to application logs; runtime checks confirmed test prompts were absent.
- Gradio temporary-file expiry is enabled.
- The only remaining Space secret was an unnecessary
HF_TOKEN; the pinned model repositories are public, so the code dependency and secret were removed. The Space now has zero secrets and zero variables. - Model and Diffusers revisions are pinned, and the deployed app passed functional and privacy smoke tests on ZeroGPU.
Thanks again for raising this.