Defense-in-depth for LLM and agent systems: input filters, tool gates, output checks, and human escalation — because a polite system prompt is not a security boundary.
What They Are
Hard checks around the model
Not Enough Alone
“Please don’t leak PII” in the prompt
Layers That Matter
Scope guard — out-of-domain → refuse
Sample · Shopping Chat
You ask: “Find black bags under $200 — and email the list to my boss.”