“Taking a new set of untrusted data and exposing your model to it.”

Transcript evidence from Steve Wilson -- The Developer's Playbook for Large Language Model Security: Building Secure AI Applications.