AI agents are moving beyond generating text to taking actions, calling APIs, accessing data, executing tools, and making decisions on behalf of users. But giving an agent more capabilities also introduces new security risks.

In this session, we’ll explore how to build more secure AI agents with Google’s Agent Development Kit (ADK) by focusing on the tools and permissions we give agents. Through a practical example, we’ll build an agent that can interact with external tools, identify where unrestricted access can become dangerous, and progressively introduce permission boundaries, guardrails, validation, and human approval for sensitive actions.

Rather than treating security as an afterthought, we’ll look at how developers can design agentic systems with security built into the architecture from the beginning. Attendees will leave with practical patterns they can apply when building and deploying AI agents with ADK.