What matters for your decision
Use least privilege, limited tools, source review and separate approval rules together. Do not assume a prompt or model filter alone can guarantee protection.
Check the boundary
- Identify which content paths enter the model and which tools could read or change records; narrow their scope before enabling them.
- Require server-side authorization and explicit confirmation for identity-bound writes; reject unexpected fields or destinations.
- Use safe synthetic test content to confirm the bot treats document instructions as untrusted and escalates when uncertain.