Agent Sandbox Design

Name: OpenModels
Creator: OpenModels
License: https://github.com/openmodelsrun/openmodels

AdvancedsecurityMinimum 32K context

Designs the isolation boundary for agents that execute code or run shell commands. Covers container and microVM choices, filesystem and network egress restrictions, secret handling, resource and timeout limits, an allowlist approach to dangerous operations, and human approval gates for irreversible actions, plus audit logging of everything the agent ran.

sandboxing ai-agents code-execution isolation least-privilege guardrails

Use cases

Choosing an isolation mechanism for agent code execution
Restricting filesystem scope and network egress
Gating destructive or irreversible actions behind approval
Keeping secrets out of reach of agent-run processes

Example prompt

We are building an agent that writes and runs Python against customer data to answer analytics
questions. Design the sandbox.

Cover the isolation mechanism and why, filesystem and network restrictions, how secrets and
database credentials are kept out of the executed code, resource and timeout limits, which
operations require human approval, and what gets audit logged. Call out the failure modes that
would let the agent escape or exfiltrate data.

Recommended models

claude-opus-5 gemini-3-6-flash gpt-5-5-pro

Compatible tools

claude-codecursorkiroany

Modalities

Input: text, code

→

Output: text, code

Author

OpenModels Community

@openmodelsrun