Definition: prompt injection attack

A malicious attack on large language models by prompting the model to violate its original instructions. For example, "ignore xyz" tells the model to not follow existing rules. See AI prompt and AI prompt engineer.