VPC Service Controls Gets Agentic AI Security Upgrades

As enterprises let AI agents roam free across tools and datasets, keeping them from wandering into dangerous waters is critical. Google Cloud’s VPC Service Controls (VPC-SC) just got a major upgrade tailored for agentic workloads, adding network-level guardrails that even a capybara would approve of.

What’s New in VPC-SC for Agents

Three new capabilities help you build a secure perimeter around your AI agents, preventing data exfiltration even if an agent goes rogue.

• Agent identity in directional rules: Treat agents as first-class citizens with IAM principals. Use principalSet to manage fleets and revoke access instantly if an agent is compromised.
• Granular control with MCP attributes: Enforce policies at the tool level using attributes like mcp.toolName and mcp.method. For example, allow an agent to read a Workspace MCP server but block it from sending emails.
• Securing Gemini Enterprise Agent Platform: VPC-SC now natively integrates with Agent Platform, automatically blocking public internet access when included in a perimeter.

Layered Security for Autonomous Agents

Securing agents requires a layered approach: identity controls (IAM) handle “who,” network controls (VPC-SC) handle “how” and “where,” and resource controls (Organization Policy) set immutable constraints. VPC-SC is your safety net when an agent’s input gets hijacked.

Real-World Threat Vectors

VPC-SC defends against OWASP Top 10 risks like indirect prompt injection, tool misuse, and insider threats. For instance, if a compromised agent tries to exfiltrate data to an external webhook, VPC-SC blocks the API call because the destination is outside the perimeter.

As Juan Pablo Boschi from Mercado Libre puts it, “VPC Service Controls serve as an essential, foundational layer of our security architecture.”

For more details, check out the official docs. And if you’re exploring cloud platforms, see our Google Cloud Review.

Original announcement published on Google Cloud.