LLM-Powered Apps (LPAs) are increasingly being deployed on shared cloud servers where multiple users interact at the same time. This creates data leak risks where one user’s private data can flow into another user’s session through shared caches, misconfigured access control policies, or concurrency bugs in multi-threaded and event-driven code. Existing taint analysis tools do not handle these risks well because they need manual source-sink marking, only track control and dataflow, and do not account for access control policies, complex app structures, or runtime environments. In this paper, we propose Access Flow Guard (AFG), a framework that automatically identifies taint sources based on user identity and permission level to detect cross-user data leaks in LPAs. Instead of manually marking sources and sinks as existing taint approaches require, AFG introduces a Multi-User, Multi-Permission (MUMP) setup. MUMP automatically identifies user input entry points by matching function signatures against patterns derived from our dataset of popular open-source LPAs. Each identified input is tagged with the user’s identity and permission level. AFG then applies Scoped Taint Pointer Analysis (STPA), building on existing pointer-based taint analysis techniques, to propagate user-tagged objects through a Pointer Assignment Graph. STPA detects where different users’ data scopes overlap. We discuss key research challenges, including treating LLM components as black boxes and refining static results through dynamic testing. This approach aims to serve as a basis for systematic security testing of LLM-integrated software in shared multi-user environments.