Safety visual guide
Tool and data trust boundaries
Where identity, policy, approval, validation, and audit controls belong around an AI tool call.
Read the visual
What this shows
A model is one participant in a larger system. The host supplies identity and policy, a tool gateway validates typed input and checks authorization, a human approves higher-impact actions, downstream systems enforce their own permissions, and audit evidence captures what actually happened.
Carry this forward
Key takeaways
- Never treat the model as the authorization boundary.
- Validate inputs before execution and results before returning them to the model.
- Use human approval for consequential actions and preserve an auditable record.