[Vertex AI,Batch] Gemini 3.1 Flash Image - GCS Authentication Error (Code 13 / impersonation_policy failure)

Gemini API
, ,
1

Model: gemini-3.1-flash-image-preview Location: global Issue Starting: April 16, 2026

Description: > I am encountering a persistent “UNAUTHENTICATED” error when using fileUri (GCS) with the Gemini 3.1 Flash Image model in Vertex AI. This issue started occurring suddenly since yesterday.

Error Log:

Plaintext

Invalid authentication from policy (go/gcs-rpc-sp): Rejected by impersonation_policy (attempt to impersonate eval-infra-scheduler-scraper@prod.google.com)
Permission 'auth.impersonation.impersonateProdUser' not granted to cloud-ml-lvm-video-server@prod.google.com

What I’ve Checked:

  1. Verified IAM roles for the Service Account (Storage Admin, Vertex AI Admin).

  2. Verified IAM roles for the Vertex AI Service Agent (Storage Object Viewer assigned).

  3. The issue persists even with 100% correct IAM settings.

This seems to be an internal Google infrastructure/policy regression regarding how the vision server fetches data from GCS. Is this a known issue?

To provide more context, I am processing image analysis with Gemini 3 Flash and image generation with Nano Banana 2 via Vertex AI batch jobs to use Nano Banana 2 for translation purposes while reducing costs.

Everything has been working fine until recently, when errors suddenly started occurring.

I am not sure if this is an internal Google issue and I should just wait, or if I fundamentally need to modify my program.

Is anyone else experiencing similar issues?

2

I have the same problem! And in his status page is all good…..

3

i also have this problem today. why?

4

Same problem here, started early this morning.

5

anyone having any luck?

6

Well you have a impersonation policy on your account. Plus can I ask what are your goals? I’m unsure of your goals because I see this policy and you’re using rpc as well…can you write out in more detail what your trying to achieve?

Thanks

7

I am currently accessing the Vertex AI Batch Prediction feature through a webpage built using the AIS build function.
I upload files to GCS, have them analyzed (Gemini 3 Flash), and then have them translated (Gemini 3.1 Flash Image).
The analysis phase is working fine, but I suspect an issue is occurring during the image-reading process during translation.
It used to work without any problems, and there shouldn’t be any issues with the service account or similar settings…

  • 4/18 The problem has been fixed now. It seems it was an internal Google issue and they fixed it.
8

What ever was fixed here - Isnt a full fix . Now the model silently drops GCS URI paths . Which is costly . So any GCS URL paths not in the same region “GLOBAL” for all gemini 3 , seems to drop .
Do we have a solution for this or are we forced to use signedURLs ?