I agree. Though even if it's running locally, if you ask it to browse the web for a specific answer it's still vulnerable

Knowing about prompt injections, I really don't like the growing integration of LLM inside real applications