Strengthening Agentic AI Security with Boundary-Based Guardrails: New Capabilities in VPC Service Controls

As enterprises scale autonomous AI agents into production environments, robust architectural guardrails that account for the nature of agents traversing tools and datasets are essential. To address this, Google Cloud has introduced new security capabilities specialized for agent workloads via VPC Service Controls (VPC-SC).

The key updates are as follows:

• Agent identity enforcement in directional rules: Agents can be treated as independent identities (IAM Principals) and added directly to the ingress and egress rules of service perimeters. Using PrincipalSet, consistent access policies can be applied to groups of agents, and if a specific agent is compromised, access can be immediately blocked at the network boundary.
• Fine-grained control using MCP attributes: As the Model Context Protocol (MCP) establishes itself as the standard integration layer for agent systems, tool-level policy enforcement has become possible. Conditional access rules can be set based on MCP attributes such as mcp.toolName, mcp.method, and mcp.tool.isReadOnly. For example, an agent can be granted read access to a Workspace MCP server while explicitly blocking it from sending emails.
• Strengthened security for the Gemini Enterprise Agent Platform: The Gemini Enterprise Agent Platform has been natively integrated with VPC-SC. By including the platform as a protected service within a VPC-SC perimeter, all public internet access to platform instances is automatically blocked without additional configuration.

Secure AI agent operations require a layered security approach that combines identity control via IAM, network control via VPC-SC and next-generation firewalls, and resource control via Organization Policy.