Source: Vertex AI — VPC Service Controls

How a VPC-SC perimeter blocks exfiltration from Vertex AI / Agent Platform, which features break inside it, and how PSC and Private Google Access fit.

Contents

Summary

Explains perimeter-based data-exfiltration protection for the platform and its operational side effects.

Key claims

  • Inside a perimeter, training data, models, online and batch inference data and Gemini calls cannot leave; access from outside is denied unless an access level allows it.
  • Producer environments lose default internet egress. Agent Runtime then needs a proxy VM with Cloud NAT for outbound calls.
  • Limitations: URL-context live fetching is disabled, request/response logging is unavailable, public 1-click Model Garden endpoints are unsupported (use private endpoints), and pipelines can’t pip-install from the internet.
  • Endpoints, and agents deployed before a project joins the perimeter, are not protected and may fail or keep internet access, so create them after joining.
  • Private Service Connect and Private Google Access give private routes but don’t remove public accessibility; VPC-SC is what blocks public access. Use the restricted VIP 199.36.153.4/30.

Feeds

Google Vertex AI

Suggest changes →