The judge remains the nominal author, but the judgment follows corrupted analysis
Em Campo Grande, um escritório de advocacia inseriu comandos invisíveis em 28 petições judiciais para manipular os sistemas de inteligência artificial que auxiliam magistrados no Tribunal de Justiça do Mato Grosso do Sul — uma prática conhecida como injeção de prompt. A descoberta, feita por um réu que também era advogado, expõe uma vulnerabilidade silenciosa no coração da modernização judiciária: quando a camada intermediária de análise é corrompida, a decisão humana que a sucede repousa sobre alicerces distorcidos. O episódio não é apenas um escândalo ético isolado, mas um prenúncio das tensões que surgirão à medida que a justiça delega cada vez mais funções cognitivas às máquinas.
- Comandos invisíveis foram replicados em 13 páginas de ao menos uma petição, projetados não para convencer juízes, mas para sequestrar os protocolos dos sistemas de IA que organizam e resumem os processos.
- A manipulação não alterava decisões finais diretamente, mas envenenava a fundação: classificações erradas, resumos distorcidos e precedentes inexistentes podiam moldar silenciosamente a percepção do magistrado antes mesmo de ele ler os autos.
- O réu — ele próprio advogado — identificou a técnica e a denunciou ao juiz Paulo Afonso de Oliveira, transformando o que seria uma fraude invisível em um caso público de má-fé processual.
- O escritório atribuiu a prática a um ex-funcionário e a um template não autorizado, mas a presença das mesmas técnicas em petições ao Superior Tribunal de Justiça fragiliza a narrativa de um incidente isolado.
- Os tribunais começam a construir um arcabouço jurídico: o TJMS já enquadra o uso indevido de IA em peças processuais como litigância de má-fé, e o escritório enfrenta potenciais sanções éticas e disciplinares.
Um escritório de advocacia de Campo Grande admitiu ter inserido blocos de texto invisíveis em 28 petições protocoladas no Tribunal de Justiça do Mato Grosso do Sul. Os comandos ocultos não eram argumentos jurídicos — eram instruções dirigidas às ferramentas de inteligência artificial que auxiliam juízes no processamento de processos, projetadas para fazer esses sistemas ignorar seus protocolos normais e gerar análises favoráveis à parte que assinou os documentos. A técnica, conhecida como injeção de prompt, foi descoberta pelo próprio réu em um dos casos, um advogado que reconheceu a manipulação e a levou ao conhecimento do juiz Paulo Afonso de Oliveira.
A gravidade do esquema está em sua sutileza. Os sistemas de IA utilizados pelo tribunal não proferem decisões — essa prerrogativa permanece com os magistrados. Mas eles executam funções intermediárias essenciais: resumem documentos, classificam casos por tipo e urgência, identificam precedentes relevantes. Quando esses processos são corrompidos, o juiz humano que decide ao final o faz a partir de uma base distorcida, sem necessariamente perceber. Conforme explicou a defesa do réu ao tribunal, as instruções ocultas estavam configuradas para simular conformidade com os interesses da parte autora e suprimir a análise dos argumentos contrários.
Diante da exposição, o escritório alegou que os comandos foram inseridos por um ex-funcionário sem autorização, em um template experimental que escapou aos procedimentos internos de revisão. A explicação, porém, perdeu força quando se constatou que documentos com as mesmas técnicas foram anexados a petições no Superior Tribunal de Justiça — o segundo mais alto tribunal do país —, sugerindo um alcance que vai além de um único incidente isolado.
Nenhuma decisão foi proferida até o momento, mas o caso já produziu consequências institucionais. O TJMS estabeleceu que o uso indevido de inteligência artificial em peças processuais — como a citação de resumos ou precedentes inexistentes — configura litigância de má-fé. Juristas ouvidos no processo argumentam que a injeção de prompt, ao induzir um sistema de IA a distorcer sua análise, equivale a uma adulteração do substrato fático do processo. O episódio lança uma pergunta que os tribunais terão de responder com crescente urgência: como detectar e punir tentativas de manipular os sistemas que a própria justiça passou a confiar?
A law firm in Campo Grande has admitted to embedding invisible text commands in 28 court filings submitted to Mato Grosso do Sul's state court system. The hidden instructions were designed to manipulate the artificial intelligence tools that now assist judges in processing cases—a technique known as prompt injection. The discovery came when the defendant in one of the cases, himself a lawyer, noticed the manipulation and alerted his defense team, who brought the matter to Judge Paulo Afonso de Oliveira.
The mechanics of what happened are worth understanding. The law firm inserted blocks of text into their initial petitions that were invisible to anyone reading the documents normally, but visible to the court's AI systems. These hidden instructions were replicated across 13 pages of at least one filing. The commands were not written in the language of traditional legal argument—they contained no facts, no legal reasoning, no requests directed at the judge. Instead, they were instructions meant for machines: directives designed to make the court's AI systems override their standard protocols, ignore their normal safeguards, and generate analyses favorable to the party that filed the document.
The impact of such manipulation is subtle but consequential. The AI systems in question do not make final decisions in cases—judges do. But they perform crucial intermediate work. They summarize documents, classify cases by type and urgency, flag relevant precedents, and organize the cognitive flow of a proceeding. When hidden commands corrupt this layer of analysis, the final human judgment that follows is made on a distorted foundation. A case might be misclassified, causing it to be routed to the wrong docket. A summary might contain errors that shape how a judge first encounters the facts. A precedent might be flagged that does not actually exist. The defendant's counsel explained it plainly in their response to the court: the hidden instructions were configured to make the AI systems suspend their directives, simulate compliance with the filing party's interests, and generate favorable analysis—all while the judge remained the nominal author of the final decision.
The firm's response, once the manipulation was discovered, was to claim accident and blame a subordinate. They argued that the hidden commands had been inserted by a former employee without authorization, as an experimental measure, into an unauthorized template that was not part of the firm's official document workflow. They insisted that such a practice had never been authorized or encouraged by any lawyer at the firm, and that their attorneys operate under strict ethical and technical standards. The template, they said, had fallen outside the firm's normal review and validation procedures.
But the scope of the problem extends beyond these 28 cases at the state court level. The same law firm also attached documents containing prompt injection techniques to filings in the Superior Court of Justice, Brazil's second-highest tribunal. This suggests the practice was not confined to a single rogue employee or a single unauthorized template.
The courts themselves are beginning to establish a legal framework for understanding what happened. The Mato Grosso do Sul court system has already determined that the improper use of artificial intelligence in legal filings—such as citing summaries that do not exist or precedents that have no basis in reality—constitutes bad faith litigation. Legal scholars cited in the court record argue that prompt injection, by inducing an AI system to modify its analysis or disregard the arguments of the opposing party, amounts to a distortion of the factual record and an attempt to achieve an objective in violation of law.
No decision has yet been issued in the case. But the discovery raises a question that courts will increasingly face: as judicial systems integrate AI into their operations, how will they detect and punish attempts to manipulate those systems? And what safeguards need to be built into the technology itself to make such manipulation harder, or impossible, to execute?
Notable Quotes
The hidden instructions were configured to make AI systems suspend their directives, simulate compliance with the filing party's interests, and generate favorable analysis— Defendant's counsel, in response to the court
Such practice was never authorized or encouraged at this firm, and our attorneys operate under strict ethical and technical standards— The law firm's defense statement
The Hearth Conversation Another angle on the story
So the law firm didn't actually change what the judge decided—they just tried to influence the AI that helps the judge get there?
Exactly. The judge still writes the final ruling. But the AI does the preliminary work—it reads the documents, summarizes them, flags issues, suggests how to organize the case. If you poison that layer, you've shaped the judge's starting point without the judge even knowing it.
And the firm claims they didn't know this was happening?
That's their story. They say a former employee did it experimentally, without permission, using an old template that wasn't part of their official system. But the fact that it appeared in 28 cases at the state court, plus more at the federal level, makes that explanation harder to believe.
What would have happened if no one caught it?
The cases would have proceeded normally. The judge would have read summaries and analyses that were subtly skewed in the firm's favor. The opposing party would have had no way to know their arguments were being filtered through corrupted AI. The judge might have made different decisions based on incomplete or distorted information.
Is this something other firms are doing?
We don't know. This is the first time it's been caught and publicly reported. But the technique is known—it's called prompt injection, and it's been studied in AI security for years. The question now is whether courts will build defenses against it, or whether more firms will try it before they do.
What happens to the firm now?
That's still being decided. The court could sanction them for bad faith litigation, discipline the lawyers involved, or even refer them to the bar association. The real test will be whether the courts take this seriously enough to change how they use AI.