Defense in Depth

Securing the Loop

You cannot rely on the LLM's built-in safety alone. You must build defenses into the orchestrator:

• Sandboxing: Run all agent code in isolated environments without network access to internal systems.
• Least Privilege: Only give the agent the exact tools it needs. Don't give a read-only agent a delete_row tool.
• Human-in-the-Loop (HITL): Require a human to click "Approve" before any irreversible action (e.g., sending an email, dropping a table).
• Input/Output Filters: Pass the agent's planned action through a smaller, fast model trained specifically to detect malicious intent before executing it.