Mobile apps are racing to ship AI assistants, copilots, and agents, but the path from the app to the model introduces a new attack surface. Prompt injection is no longer just a web chatbot problem: attacker-controlled data can reach an LLM through API responses, local storage, runtime manipulation, retrieved content, and other context the app trusts. In this session, we'll threat-model the mobile LLM gateway and follow real attack paths from Android into the model's context. We'll examine direct and indirect prompt injection, runtime manipulation, and why traditional mobile security controls alone aren't enough. You'll leave with practical techniques for identifying injection paths, defining trust boundaries, and protecting prompts before they reach the model.