Google Cloud’s Vertex AI Hit by Vulnerability Enabling Sensitive Data Access

Unit 42 disclosed a critical flaw in Google Cloud's Vertex AI Agent Engine where default service-agent permissions and the use of pickled AI agents allow an attacker to retrieve P4SA credentials from the metadata service and escalate privileges across GCP, potentially exposing Cloud Storage buckets, Artifact Registry repositories, deployment files, and Google Workspace data; Google worked with Unit 42 to address the issue and recommends using BYOSA and strict least-privilege controls for AI agent deployments.