Back to OWASP Top 10#4 High Risk

LLM04:2025 Data and Model Poisoning

Manipulation of training data or fine-tuning processes to introduce vulnerabilities, backdoors, or biases that compromise model integrity and security.

Vulnerability Overview

Data and Model Poisoning occurs when training data, fine-tuning data, or feedback mechanisms are manipulated to introduce vulnerabilities, backdoors, or biases into LLM applications.

Impact Level

High

Attack Vector

Training Data

Exploitability

Medium

Pre-training Data Poisoning

Manipulation of the foundational training dataset used to train the base model.

  • • Large-scale dataset contamination
  • • Backdoor trigger insertion
  • • Bias amplification
  • • Misinformation injection
Fine-tuning Poisoning

Manipulation of fine-tuning datasets or processes to alter model behavior for specific tasks.

  • • Task-specific manipulation
  • • Adversarial examples
  • • Instruction following corruption
  • • Safety alignment bypass