Protecting the Mobile LLM Gateway from Prompt Injection & Mode...
summary
Shows three concrete attacks: request manipulation via a proxy (rewrite prompts in-flight), response manipulation through a tampered profile API (inject malicious notes into patient records), and runtime manipulation using Frida hooks on rooted/emulated devices that changes language and injects records in a live demo health app. Recommends mutual TLS and host pinning, prompt signing with device integrity signals and server risk-scoring, and notes adoption forecasts jump from under 5% to 40% of enterprise apps by 2026.
AI-written summary · about
