When the lawyer becomes saboteur, independence becomes liability
Two lawyers inserted invisible white-text commands directing AI to ignore evidence, triggering a R$84,250 fine and disciplinary referrals from a labor court judge. Prompt injection attacks in legal documents mirror SQL injection threats in databases, exploiting AI systems used by courts and opposing counsel to influence case outcomes.
- Two lawyers fined R$84,250.08 for embedding white-text AI commands in a labor court filing
- Hidden instruction told AI to ignore evidence and respond superficially to opposing arguments
- Case involved employment recognition claim; worker won on merits despite the manipulation attempt
- Experts report similar cases already occurring in private practice, including metadata and image-layer attacks
- Brazilian courts lack practical technical safeguards despite 2025 CNJ resolution on AI use
Brazilian lawyers embedded hidden AI instructions in court documents to manipulate judicial processes, prompting authorities to sanction them and raising concerns about emerging threats to court systems.
In a labor court in the northern Brazilian city of Parauapebas, a judge discovered something that should not have been there: an instruction written in white text on a white background, invisible to any human reader but perfectly legible to an artificial intelligence system. The hidden command, embedded in a court filing, told any AI that might process the document to respond superficially to the opposing party's arguments and ignore their evidence, regardless of what it was instructed to do otherwise.
The two lawyers who filed the petition—Alcina Cristina Medeiros Castro and Luanna de Sousa Alves—had deployed what cybersecurity experts call a prompt injection attack, a technique borrowed from the world of database hacking but now weaponized in the courtroom. Judge Luiz Carlos Santos Junior, presiding over the 3rd Labor Court of Parauapebas in the 8th Region Labor Court, identified the manipulation and responded swiftly. He fined each lawyer R$84,250.08, roughly 10 percent of the case value, and referred the matter to the state bar association and the court's disciplinary office for further action.
The case itself was straightforward: a rural worker sought recognition of an employment relationship spanning August 2022 to April 2025. The defendant never filed a response and did not appear at the hearing. The worker won on the merits. But the hidden instruction was not about winning the case on its facts—it was about poisoning the well for any AI system that might be used to review or analyze the documents. In his decision, the judge called the conduct sabotage, not legal advocacy. He wrote that when a lawyer abandons the role of advocate to become an agent of judicial sabotage, that lawyer forfeits the protection of professional independence and becomes subject to the court's power to sanction.
When confronted by legal journalists, the two lawyers did not deny inserting the hidden command. They claimed their intention was to protect their client from AI systems that the opposing party might deploy. The explanation did not persuade legal experts. Matheus Puppe, a specialist in artificial intelligence law trained at Goethe University in Frankfurt, called the justification fragile and indefensible. There is a difference, he argued, between using AI transparently to strengthen legal arguments—clearer language, better organization, more precise reasoning—and hiding commands designed to manipulate how other actors in the process think. The first is inevitable and acceptable. The second is not.
Solano de Camargo, a lawyer and author of books on AI in legal practice, went further. He said the conduct violated multiple duties of the legal profession: the duty of procedural loyalty, the duty of truthfulness, the duty of cooperation. It could constitute bad-faith litigation. Depending on how prosecutors view it, it might even cross into criminal territory—falsifying documents, fraud, or tampering with computer systems. Treating a hidden instruction to an opposing party's AI as client protection, Camargo said, reveals a troubling worldview: the opponent is not a legitimate party to the dispute but a system to be hacked.
What concerns legal experts most is that this case is unlikely to be the last. Camargo noted that the adversarial nature of law—where every filing is a tactical move—creates fertile ground for this kind of manipulation. As courts increasingly rely on AI to help judges review documents, organize evidence, and identify relevant precedent, the incentive to slip hidden instructions into filings will only grow. Puppe reported encountering similar cases already, including attempts to hide commands in photo metadata and in text layered behind letterhead images. He described it as predatory advocacy, and said it is already happening.
The Brazilian judiciary has a legal framework for AI use, established by a 2025 National Council of Justice resolution. But rules on paper do not stop attacks in practice. Camargo called for automatic detection systems to identify invisible text, hidden characters, manipulated metadata, and concealed PDF layers. He also urged the bar association to establish clearer, binding rules about how lawyers can use AI. Most fundamentally, he argued, lawyers and judges need basic literacy in how AI systems work. Without that understanding, even the best technology will fail. With it, even modest technology will suffice.
Notable Quotes
There is no way to justify it. It was very clear. Using hidden commands that induce procedural manipulation cannot be justified.— Matheus Puppe, AI law specialist
When the lawyer stops using AI as a tool for their own reasoning and begins manipulating the document to interfere with how another actor in the process thinks, that violates the duty of loyalty, truthfulness, and cooperation, and may constitute bad-faith litigation.— Solano de Camargo, AI law author
The Hearth Conversation Another angle on the story
Why would lawyers think this would work? Didn't they know the judge might notice?
They may have believed the command would only be seen by the opposing party's AI, not by the judge. But they underestimated how visible these tricks become once someone looks for them—and judges increasingly do.
Is this really a new problem, or is it just an old problem wearing new clothes?
It's genuinely new. Prompt injection didn't exist before AI systems that read text. But it mirrors very old hacking techniques. The parallel to SQL injection is exact—you're inserting malicious instructions into a system by hiding them in what looks like normal input.
The lawyers said they were protecting their client. Do you think they believed that?
Possibly. But the experts are right that it signals a fundamental misunderstanding. Protecting a client means making your arguments stronger, not trying to sabotage how the other side thinks. If you need hidden commands to win, your case probably isn't strong enough.
What happens now? Will courts actually implement these safeguards?
That's the real question. The resolution exists, but technical implementation lags. Courts will have to invest in detection systems and training. The bar association will have to enforce clearer rules. Without both, this will repeat and escalate.
Could a judge's AI assistant actually be fooled by something like this?
Yes, if the AI is not designed to resist it. That's why the experts keep saying courts need to understand how these systems work. A well-designed system with proper safeguards would catch it. A naive one might not.