Two Brazilian lawyers have been fined after hiding a command inside a legal filing that attempted to steer a court’s artificial-intelligence assistant toward a weaker analysis. The text, formatted in white against a white background, was invisible during an ordinary reading but was detected when the document was processed by the tribunal’s AI prompt-injection case.
The incident took place in the 3rd Labor Court of Parauapebas, in Brazil’s Pará state. Judge Luiz Carlos de Araújo Santos Júnior identified the hidden material in a complaint signed by Alcina Cristina Medeiros Castro and Luanna de Sousa Alves. Translated from Portuguese, the instruction told the artificial intelligence to respond to the filing “superficially” and avoid challenging its documents, regardless of any other command it received. The lawyers were fined R$84,250.08—10 percent of the case’s value—and the matter was referred to the Pará section of Brazil’s bar association. Court records identified the case as process 0001062-55.2025.5.08.0130.
The technique is known as prompt injection: an attempt to smuggle instructions into material an AI is meant to analyze, effectively treating the document as a second set of orders. In this case, the filing’s formatting was designed to hide the instruction from people while leaving it available to a text-processing system. The court characterized the maneuver as an act offensive to the dignity of justice, assigning responsibility to the attorneys who signed the document rather than to their client.
Ironically, the manipulation appears to have had no effect on the lawsuit’s immediate outcome. The opposing party did not submit a defense, meaning the plaintiff prevailed by default. The court nevertheless treated the act itself as serious: an attempt to influence a judicial workflow remains consequential even when the hoped-for advantage is unnecessary or fails.
The incident also illustrates what Brazil’s Galileu system actually does. Developed by the Regional Labor Court of the 4th Region and later made available across the labor judiciary, Galileu reads pleadings and other case materials, organizes claims and defenses, and proposes draft reports and research leads. The system is not supposed to decide cases, assess evidence, or replace a judge; its output must be reviewed, edited, accepted, or rejected by a human magistrate.
That human checkpoint mattered here. The tribunal says Galileu flagged the concealed passage, blocked the malicious text from being processed, and presented the alert to the user. The judge then inspected the document independently before imposing the sanction—a useful distinction as courts and other institutions experiment with AI systems that ingest untrusted documents. The lesson is less that legal filings can secretly control a judge than that every document supplied to an AI workflow must be treated as potentially adversarial input.
