- Definition: users make the app do things it wasn't built for. LLMs are instruction-followers, so without careful design any capable model behind your app can be repurposed.
0:30 - Real case: a user asked Klarna's customer-support chatbot to help with Python coding — and it complied.
0:30 - Why it matters: not directly harmful in itself, but the app gets hijacked for tasks that waste resources (you pay for the tokens) and confuse users.
0:48
Mitigation covered later: a topic classifier in front of the LLM (product design choices).