Detecting Data Leaks in Multi-User LLM Apps via Automated User-Scoped Taint Analysis
summary
Finds a real Open WebUI bug that leaks one user's PDF-derived embeddings into another user's session and measures that 42% of popular open-source LLM apps support multi-user mode. Implements Access Flow Guard (AFG) that auto-identifies user-scoped sources via a function-signature dataset (OpenAI, Anthropic, Gemini), tags inputs with permissions in a Multi-User, Multi-Permission (MUMP) setup, and propagates tags with Scoped Taint Pointer Analysis (STPA). Prototype reports true positives and zero false positives on the evaluated apps.
AI-written summary · about
chapters (8)
- 0:00Detecting LLM data leaks
- 2:10Open WebUI leak bug
- 5:22Limitations of taint tools
- 8:50User-scoped taint analysis
- 12:28Cache-based data leak example
- 14:36MUMP and STPA for scope overlaps
- 16:43Prototype results and future work
- 17:45Q&A
related talks
49mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
46mGuarding the LLM Galaxy: Security, Privacy, and Guardrails in the...
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
46mBreaching LLM-Powered Apps: Overcoming Security and Privacy Challenges
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
53mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
44mI Abuse Your LLM Live with 10 Security Vulnerabilities