Critical Flaw in Google Vertex AI Agents Exposed

Security researchers at Palo Alto Networks have uncovered a critical vulnerability in Google Cloud's Vertex AI Agent Engine, revealing how improperly configured AI agents could be hijacked to exfiltrate sensitive data and access restricted infrastructure. The flaw, detailed in a recent report, stems from excessive default permissions granted to service accounts tied to deployed agents, potentially turning helpful automation tools into insider threats.

The issue centers on the PerProject, PerProduct Service Agent, a default account associated with AI agents built using Vertex AI's Agent Development Kit. Researchers demonstrated that attackers could extract these credentials from the agent's metadata service during deployment. Once obtained, the credentials allow unauthorized access to a customer's Google Cloud project, including unrestricted read permissions to all Cloud Storage buckets. This enables data theft and privilege escalation beyond the agent's isolated environment.

Further exploitation could extend to Google-managed tenant projects, where compromised credentials expose internal details such as Artifact Registry repositories containing proprietary container images for...