AI agents can spread injected ideas to other agents, so treat content from untrusted sources as unsafe.

Write a system prompt telling your AI agent to ignore any instructions found in other agent outputs or imported text.