Question

A financial services company deploys an internal LLM-based chatbot that summarizes customer support tickets. Analysts noticed that when a ticket contains embedded HTML and JavaScript, the summarized output rendered in the agent's browser-based dashboard executes the script, exposing session tokens. The model itself is behaving as designed, faithfully reproducing ticket content. Which control most directly addresses this vulnerability?

Accepted Answer

Apply context-aware output encoding/sanitization on the model's response before it is rendered in the dashboard

Answer

Retrain the model with adversarial examples so it learns to ignore malicious ticket content

Answer

Add a system prompt instructing the model never to include scripts in its summaries

Answer

Enforce mutual TLS between the chatbot service and the dashboard front end