Lawyers hid a secret AI command in a Brazilian court filing. It backfired.
A labor court's AI tool spotted invisible text ordering it to go easy on a case, and the judge fined the two lawyers who signed it roughly $16,000.
What happened: A labor court in Brazil's Pará state caught two lawyers trying to secretly manipulate its AI review tool. The court, part of the 8th Regional Labor Court, uses an AI system called Galileu to help screen case filings. Buried inside a routine petition was a block of text in white font on a white background, invisible to anyone reading the document on screen or in print, but still readable by software. The hidden message instructed any AI reading the file to review the petition only superficially and avoid challenging the attached documents, no matter what other instructions it received later. The judge spotted the trick, fined the two signing lawyers 10% of the case's claimed value, reported as about $16,000, for contempt of court, and referred the matter to Para's bar association for possible discipline.
Why it matters: Courts in Brazil and elsewhere are increasingly using AI to triage, summarize, or pre-review the flood of filings judges face. That only works if the AI's output reflects a case's real merits rather than whatever a party managed to smuggle past a human reader. This case shows lawyers are already testing whether hidden instructions can quietly tilt an AI's assessment in their favor, effectively arguing their case to the machine instead of the judge. If it had worked, a court could have been pushed toward accepting weak or disputed evidence without real scrutiny, undermining the other side's right to a fair hearing. Catching it this time doesn't settle how many subtler attempts might already be slipping through elsewhere.
How the trick worked: The technique is called prompt injection: hiding instructions meant for an AI inside content a human reader won't notice, hoping the AI follows them as if they were legitimate commands. Here, the lawyers reportedly used ordinary-looking white text on a white page background, invisible to the eye but fully readable by any software, including an AI, processing the document's underlying text. The hidden line, translated from Portuguese, told the AI to treat the filing gently and to ignore any later instruction telling it otherwise. Exactly how Galileu detected the buried text instead of obeying it hasn't been detailed publicly, but the court's ruling confirms the system flagged the attempt rather than acting on it.
What happens next: The judge fined the two lawyers who signed the petition and ordered the case reported to the Para branch of Brazil's bar association, which could pursue its own professional discipline, in theory including disbarment. The ruling is now circulating among Brazilian legal commentators as an early test case for how courts should police AI misuse by the people appearing before them. Online discussion has already raised a harder question the case doesn't answer: the hidden text here was large enough to get caught, but a smaller, subtler version might not be. That leaves open whether courts need routine, systematic checks for hidden text in every filing, rather than relying on an AI tool to notice each injection attempt on its own.