As organizations move from static chatbots to autonomous AI Agents, the attack surface shifts from simple user inputs to complex, multi-step execution environments. To deploy these agents safely, a "perimeter-only" defense is no longer sufficient. We need to protect in multiple layers of the Networks and a Zero Trust architecture that treats every agent action from model inference to tool execution as a potential risk.

In this session, we will architect a comprehensive security posture for AI agents using Google Cloud’s advanced networking and security stack. We will move beyond basic IAM to explore how Model Armor intercepts prompt injections, how VPC Service Controls (VPC-SC) create cryptographic boundaries around sensitive data, and how NGFW Enterprise and Secure Web Proxy provide deep packet inspection and egress control for agentic workflows.

What You Will Learn
The Anatomy of Agentic Risk: Understanding prompt injection, data exfiltration, and "confused deputy" scenarios in autonomous systems.

Inbound Integrity with Model Armor: How to sanitize inputs and redacting sensitive PII before it reaches your LLM.

Data Sovereignty with VPC-SC: Building a "Service Perimeter" that prevents agents from moving data across unauthorized projects.

Egress Governance: Using Secure Web Proxy and NGFW Enterprise to ensure agents only communicate with "Known Good" APIs and domains and targets.

Unified Posture Management: Integrating these controls into Security Command Center (SCC) for automated enforcement and threat detection.

Target Audience
Cloud Architects, Security Engineers, and AI/ML Developers looking to move agentic workloads from "experimental" to "enterprise-ready."