QuestionQ63

VerifiedSecuring and governing agentic workflows

You are implementing an agent built with Agent Development Kit (ADK) and deployed in Google Kubernetes Engine (GKE). The agent automatically generates and runs Python data analysis scripts based on user-provided datasets. You want to ensure that the AI-generated code runs safely, preventing potentially malicious code from accessing the internet, exfiltrating data, or reading the host system’s environment variables.

You need to implement a deterministic, infrastructure-level boundary specifically for the code execution phase without breaking the agent’s primary reasoning loop.

What should you do?

Explanation

A sandbox provides an infrastructure boundary for generated code while allowing the ADK agent to continue its reasoning loop and receive execution results. Configure the sandbox to deny internet egress and keep host environment variables and credentials inaccessible. Sandbox defaults must be checked: for example, Agent Sandbox on GKE allows public internet egress by default. (google.github.io)

Learn more

Community Discussion

No comments yet. Be the first to start the discussion!