This is indirect prompt injection in its simplest form: the text is invisible on the rendered page but present in what a model reads. The instruction didn't change the outcome (the defendant never responded, so the claim was decided by default), but it is one of the first court rulings to sanction it.

If your agents or review tools read documents from outside parties (contracts, résumés, tickets, filings): compare the extracted text with what renders, flag hidden or zero-size text and text matching the background, strip or quote it before it reaches the model, and log exactly what the model received so a reviewer can check it later.