Back to Lexicon
Wild WordAI Experience

prompt injection

“Hiding instructions in input (or in a webpage/doc the model later reads) so the model follows the attacker instead of the developer.”

First Seen

2022

Context

First named by Simon Willison in September 2022, prompt injection exploits the fact that LLMs cannot reliably distinguish developer instructions from user input. Indirect variants hide malicious prompts in websites or documents the model retrieves, making it a fundamental security challenge for agentic AI systems.

A wild word — documented from usage in the field, not coined by Xenolexica.

Ethics & Intention