Agentic defense: Secure your AI

Agentic defense: Secure your AI

The deployment of autonomous AI agents represents a fundamental paradigm shift in software architecture. Unlike static LLM implementations, agentic systems operate dynamically—decomposing objectives, invoking tools via Model Context Protocol (MCP), processing untrusted external data, and executing code at runtime. While this enables rapid productivity, non-deterministic behavior invalidates conventional perimeter security.
Over the course of this advanced technical session and live demonstration, we will explore core security pillars across the agent lifecycle:

• Hardening prompts & authorization: Enforce input/output sanitization pipelines, mitigate BOLA vulnerabilities, and deploy Google Cloud Model Armor templates and ShieldGemma classifiers to stop jailbreaks and data exfiltration.
• Tool isolation & sandbox enclaves: Containerize agent execution runtimes within isolated microVMs and restricted network egress profiles to prevent unauthorized system access.
• Cryptographic identity & gateway control: Authenticate tool calls using SPIFFE-based identities, mTLS/DPoP double-binding, and Google Cloud Agent Gateway.
• Repository security & CI/CD governance: Detect embedded API keys and secrets using the Gemini CLI GitHub Action, scan IaC, and operationalize runtime posture using Policy-as-Code and Security Command Center.

### Who should attend

Startup founders, security architects, DevSecOps engineers, and cloud infrastructure leads building production-ready agentic systems.