Breaching LLM-Powered Apps: Overcoming Security and Privacy Challenges
summary
Live demo overwrites the app’s terms of use via a path traversal upload and RAG poisoning to add an exception that triggers a $100 credit and cancels booking ABC123, using Burp Suite to post the exploit; a SQL injection then injects fabricated chat memory so the assistant confirms cancellation. The presenter shows GPT‑3.5 is easy to gaslight while GPT‑4 resists, and implements an input guardrail (LangChain for Java) plus LLM‑as‑judge sanitizers, restricted tool permissions and human‑in‑loop checks as practical mitigations.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 3:59LLMs like children with toys
- 8:37LLMs context and memory
- 10:19RAG poisoning rental app
- 18:01Chat memory SQL injection
- 24:17LLM permissions MCP servers
- 27:27Prompt injection risks
- 32:26Input guardrails mitigation
- 36:58Tool calling local models
- 40:11LLM security remediation
- 43:00Data privacy model selection
- 45:05General rules summary
related talks
44mSecuring LLM-Powered Apps: Overcoming Security and Privacy Challenges
53mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
51mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
49mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
45mSecuring LLM Applications: Overcoming Security & Privacy Challenges
20mUnderstanding and Mitigating Risks in LLM-Powered Web Apps