Definition
Data poisoning
In one sentence
Data poisoning is an attack that corrupts the data used to train or inform an AI model, causing it to behave incorrectly, introduce hidden backdoors or produce manipulated outputs.
Data Poisoning Explained
AI models learn from data. If an attacker can insert, modify or remove some of that data, they can influence how the model behaves. Poisoning can target the original training data, fine-tuning data, or the documents and knowledge bases that AI assistants retrieve answers from, an approach often called retrieval-augmented generation.
For most small businesses, the realistic exposure is not training large models but connecting AI tools to internal content. If an AI assistant answers staff questions from a shared document library, anyone who can edit those documents can shape its answers. Malicious or simply wrong content can then spread with the apparent authority of the AI.
Controls include restricting who can modify data sources used by AI tools, keeping provenance and version history, reviewing AI outputs used for important decisions, and choosing vendors who explain how they protect model training pipelines. Data poisoning is closely related to prompt injection, and both belong in an organization's broader AI governance.
General information for Canadian organizations, not legal advice.