Sandbox Your Coding Agents

Every week there’s a new story on how an AI agent has deleted stuff it shouldn’t have: local files, cloud documents, backups etc. It feels like DOS all over again. And the news is that you can run coding agents in a safer manner, in a sandbox. Here are the main techniques, and watch for more details:

• run on a separate machine with a limited user and specific firewall rules
• mediate the agent’s calls through scripts that can only do the safe actions
• run in docker, mount only the project folder
• mount a tmpfs over the folders you don’t want the agent to access
• extract all your secrets from the code (need we say it?)
• add a second docker container that serves as network router to limit the actions of the agent; keep the agent container in a virtual network with the router container
• the router container can limit urls, ports, and services