Skip to content
Book Open access

Janus: Characterizing Bidirectional Privacy and Security Risks in LLM-Integrated Android Apps

Nov 2026 · 0 citations

Abstract

The rapid integration of large language models (LLMs) into mobile applications has introduced a new class of privacy and security risks that existing analysis techniques are poorly equipped to address. Unlike conventional third-party service integrations, LLM integrations are bidirectional: sensitive user data flows into model endpoints as part of carefully constructed prompts, and model-generated text flows out into app logic, UI surfaces, notifications, and downstream APIs. Both directions introduce distinct risks, including data leakage, unsafe reuse of model outputs, and input-safety weaknesses, that have not been systematically studied at scale. We present Janus, a static analysis of end-to-end LLM interaction pipelines in Android apps. It recovers request construction semantics, identifies response-entry points and their downstream propagation, and employs a multi-agent reasoning layer to classify privacy- and security-relevant findings. We evaluate Janus on a curated micro-benchmark suite and use it to conduct the first systematic characterization of privacy and security risks in LLM-integrated Android apps. We find that such risks are prevalent in practice, with recurring categories such as client-side secret exposure (e.g., embedded API keys and system prompts), sensitive-data outsourcing to third-party LLM services without corresponding disclosure, secondary leakage of LLM responses to downstream components, and prompt/input-safety weaknesses such as prompt-injection-prone handling. We also present representative case studies that illustrate how these risks manifest in deployed apps and what concrete harm they may enable. In line with our ethical process, we responsibly disclosed critical findings to affected developers.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.