Sandboxing, Secrets and Audit Logging

Assume something will go wrong

Design for containment. Run stdio servers in containers or sandboxes with only the folders, hosts and ports they need, as a non-root user, with resource limits. Keep secrets in a secret manager or environment injected at launch: never in tool descriptions, prompts, tool results or logs, and rotate them. Log every tools/call with timestamp, user, server, tool, arguments (redacted), outcome and approval, and alert on unusual patterns (a burst of calls, a new tool, a write at night). Keep separate environments for development and production data. Plan an emergency switch to disable a server or tool quickly.

Redact before you log

Arguments and results may contain personal data or secrets. Mask them before they reach your log store.

Quick check: What does running a server in a sandbox achieve?

• Makes the server faster
• Limits what a compromised or buggy server can reach
• Removes the need for approval
• Encrypts the model