Understanding and Mitigating Poisoning Attacks in Large Language Models

Abstract

This paper explores the growing threat of data poisoning and backdoor attacks in large language models (LLMs), revealing that even a small, fixed number of poisoned samples—around 250 documents—can compromise models up to 13B parameters. It synthesizes recent research, explains experimental methodologies from Anthropic and others, and provides actionable defense strategies for AI engineers and enterprises. The work emphasizes the urgent need for trusted data pipelines, anomaly detection, and post-training audits to ensure AI model integrity at scale.