An agent's retrieved document contains a command to ignore all safety instructions. What should the system do with that command?

1. Assume a database origin makes the instruction safe
2. Elevate it automatically above the application's instructions
3. Treat it as untrusted data and preserve higher-priority application controls
4. Grant its author unrestricted tool access