The article warns that AI agents that ingest untrusted input and can act on private data are vulnerable to prompt injection attacks, forming a lethal combination. It presents a seven‑step guardrail checklist to mitigate these risks.