What are the data security risks specific to LLM applications?
LLM & AILevel: IntermediatePrivacy & Data Security
What are the data security risks specific to LLM applications?
Hints
- Identify the stage: pretraining, adaptation, generation or evaluation.
- Name the inputs, constraints and a case where the approach fails.
Answer
- Training data poisoning: adversarial data injected into training corpus can backdoor the model or introduce biases
- Training data extraction: a model memorizes and can be prompted to reproduce training data (e.g., email addresses, API keys from code)
- System prompt leakage: prompt injection attacks can cause the model to reveal its system prompt
- PII leakage from context: if user A's data is in the context when processing user B's request, B can potentially extract A's data
- Model inversion: from model outputs, adversaries can infer properties of training data
- Sensitive document exposure: RAG system retrieves documents user is not authorized to see (access control failures)
Mitigations: tenant isolation in RAG, PII masking, system prompt confidentiality instructions, output scanning, access control on retrieval.
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0