Skip to content
Quantoros

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

  1. Training data poisoning: adversarial data injected into training corpus can backdoor the model or introduce biases
  2. Training data extraction: a model memorizes and can be prompted to reproduce training data (e.g., email addresses, API keys from code)
  3. System prompt leakage: prompt injection attacks can cause the model to reveal its system prompt
  4. 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
  5. Model inversion: from model outputs, adversaries can infer properties of training data
  6. 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

Related questions