Plaintiff busted trying to use AI prompt injection to win court case, hides text instruction in filing — demands AI model reviewing the text should side with him, rumbled because of strange white spaces in text
TL;DR
A self-represented plaintiff in a Connecticut court added a hidden AI prompt injection attack in their filing in a failed attempt to influence a decision. The court bars them from submitting documents electronically and instead must print them and hand to the clerk of court.
Nauti's Take
Teams using AI for legal or administrative workflows should treat every incoming document as untrusted input. Before testing such a workflow, inspect visible and invisible characters, formatting, embedded instructions, and the separation between document content and system instructions, especially when the model can influence a decision.
Sources