Brazilian lawyers busted for attempting to manipulate court's AI with prompt injection hidden in filing
TL;DR
The lawyers of a plaintiff in a labor case attempted to use prompt injection to influence the court's AI but were caught. They won the case anyway, as the defendant did not respond, making the move totally unnecessary. The lawyers of a plaintiff in a labor case attempted to use prompt injection to influence the court's AI but were caught. They won the case anyway, as the defendant did not respond, making the move totally unnecessary.
Nauti's Take
Teams using AI with documents should add prompt injection to every practical test: instructions found in PDFs, legal filings, and emails need clear separation from system guidance, sanitization, and logging. The first check is whether the model follows hostile document instructions and what human review happens before an AI-assisted decision is accepted.
Sources
