CVE-2026-2473

Vertex AI Experiments could save model artifacts to storage controlled by another tenant

Google Cloud · Vertex AI Experiments

First documented disclosure · Review checkpoint

Summary

The storage location used by Vertex AI Experiments could belong to someone other than the owner of the model.

Severity and exploitation evidence

High · 7.7/10 · CVSS 4.0. Severity source.

Exploitation not established: Google documented a vulnerability and a correction. The reviewed sources identify no exploitation against customer environments.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

Automatic storage selection relied on a predictable bucket name without establishing ownership.

Conditions required

A model workflow used automatic storage selection and a third party already controlled the expected bucket.

Potential impact

Google lists possible model theft, model poisoning and cross-tenant code execution.

Affected versions and configurations

Google identifies 1.21.0 through versions before 1.133.0.

Fix and mitigation status

Google identifies 1.133.0 and later as containing the mitigation.

Practical next steps

Google says the mitigation is applied and no customer action is needed. As an additional configuration check, verify pinned SDK versions and storage ownership.

Operating lesson

Check storage ownership before writing or loading model artifacts.

Evidence and limits

A January 7 change required explicit storage configuration and shipped January 8. Those releases did not describe the security impact. February 20 is the earliest verified public disclosure of the vulnerability as such.

Sources