Secure Meta Llama in a customer onboarding assistant

Question

You are building a customer onboarding assistant on a self-hosted Meta Llama model (Llama 3.1 or later, with tool calling). It answers product questions, reads customer-uploaded documents, and can call tools to update the customer's profile. Security review flagged prompt injection, data leakage, and unauthorized tool access. How would you assess and mitigate these risks using both general controls and Meta's Llama safety tools?

Accepted Answer

I would map what the assistant can read and do, then put the real controls outside the model: every tool call is validated and authorized by the backend for the current customer, uploaded documents are treated as untrusted data, and retrieval only returns that customer's records. On top of that, Meta's safety models help: Prompt Guard to detect injection and jailbreak attempts in inputs and documents, and Llama Guard to classify unsafe inputs and outputs. Finally I would red-team it with injection and cross-customer tests and track blocked attempts and false refusals.

Secure Meta Llama in a customer onboarding assistant

The interview question

Deep explanation

Simple explanation