Training Data Poisoning

In simple terms

If bad examples enter a learning set, the model can absorb the pattern before anyone asks it a question. The effect may be broad or limited to a trigger.

Technical explanation

An adversary may insert mislabeled records, repeated false claims, or trigger-bearing samples into pre-training, fine-tuning, or embedding data. Dataset lineage and validation matter because the error can persist after training.

Poisoning targets examples used to train or fine-tune a model, including labels, text, or repeated patterns. It is harder to diagnose after training, making dataset lineage and review important.

Why it matters

Reviewing data sources and training changes is as important as checking prompts at runtime.