Secure code execution

A CodeAgent executes model-generated Python. Running that on your own machine is risky, so smolagents ships several sandboxes.

Local Python interpreter (default)

By default code runs in a restricted interpreter that blocks dangerous builtins and only allows imports you explicitly whitelist:

agent = CodeAgent(
    tools=[...],
    model=model,
    additional_authorized_imports=["numpy", "json"],
)

This is safer than raw exec, but not a hard security boundary — untrusted tasks should use a real sandbox.

E2B — cloud sandbox

agent = CodeAgent(tools=[...], model=model, executor_type="e2b")

Docker — local sandbox

agent = CodeAgent(tools=[...], model=model, executor_type="docker")

Each agent step runs inside an isolated container, so even if the model writes destructive code it cannot touch your host. Use a sandbox whenever the agent handles untrusted input.