Data Poisoning attacks on Enterprise LLM applications

### Executive summary: This report describes the threat of data poisoning against Large Language Models (LLMs), explaining how manipulated training data or malicious inputs (prompt injection, poisoned models, unsanitized scraped HTML) can subtly corrupt model behavior and outputs. It gives illustrative incidents, highlights detection challenges, and recommends mitigations including rigorous data validation, red teaming, layered security, and automated model scans.