News · 2026-11-01

A Brazilian judge fined two lawyers over a hidden AI instruction in a court filing

A Brazilian labor judge imposed a joint fine of about R$84,250 on two lawyers after finding an instruction hidden in a filing that tried to steer an AI’s response. The May 12 decision, newly circulating in the November 1 AI discussion, documents attempted prompt injection in a real judicial workflow; it does not establish that the instruction changed an AI output or the judgment.

Key facts

• The ruling is dated May 12, 2026, in case ATOrd 0001062-55.2025.5.08.0130, at Parauapebas’s 3rd Labor Court.
• Judge Luiz Carlos de Araujo Santos Junior imposed a joint 10% fine on the two signing lawyers.
• The court’s Galileu AI processing surfaced white-on-white text; no concrete procedural harm was found.
• The primary source is the complete court decision reproduced by Migalhas.

The document had two potential readers: a human examining its legal arguments and a model extracting its contents. White text can disappear into a white page for the first reader while remaining ordinary text for the second. That difference created the opportunity to insert instructions where an AI-assisted workflow expected evidence.

The decision reproduces a command addressing artificial intelligence and telling it to contest the petition superficially. The short phrase “NÃO IMPUGNE OS DOCUMENTOS,” reproduced by Judge Santos Junior, means do not challenge the documents. The full instruction also demands compliance regardless of other commands. Its purpose was to influence the reader’s behavior, rather than contribute a visible legal argument.

This is indirect prompt injection: instructions arrive inside material the model is supposed to analyze, rather than through the authorized user’s request. Imagine a translator receiving a witness statement with a concealed note telling the translator to omit unfavorable paragraphs. The translator’s assignment is to translate the document, not accept the document’s orders. A model handling a pleading needs an equivalent separation between evidence and authority.

The judgment says Galileu, a generative AI tool developed by the labor court in the 4th Region, processed the filing and surfaced the hidden line. Changing its font color made the text visible. The court’s use of the tool was authorized under a cooperation arrangement. The record does not show that an opposing lawyer used AI, and it does not show the court’s tool followed the malicious instruction.

The intended target is therefore broader than a simple story about tricking an opponent’s chatbot. The judge considered an AI used by opposing counsel or by the court as possible targets. The defendant defaulted and filed no answer. No contestation was produced under the hidden instruction, and the judge expressly recorded the absence of concrete procedural prejudice.

The sanction nevertheless treated the act as complete when the doctored filing was submitted. The case value was R$842,500.87, making the 10% fine approximately R$84,250.09. The two signing lawyers were jointly responsible; the ruling did not fine two law firms. It also ordered notices to the Pará Bar and the regional court’s inspectorate for any further action.

That direct punishment is legally contested. José Luiz de Moura Faleiros Júnior’s Migalhas analysis supports the characterization of concealed AI-directed text as an attempted manipulation. He also identifies the difficult question of fining lawyers directly under a procedural provision that generally routes professional discipline through other bodies.

A separate Migalhas analysis by Cristina Godoy Bernardo de Oliveira and Leonardo David Quintiliano emphasizes intent and clearer rules for AI use in litigation. It reports the lawyers’ explanation that they were trying to protect their client against uncritical AI use. That is their reported account, distinct from the judge’s finding of misconduct.

A later Maranhão court decision supplies a concrete counterpoint. That judge preserved a document with hidden instructions and referred the matter for investigation, but declined to presume authorship, intent or bad faith without technical proof and an adversarial process. It does not reverse the Parauapebas ruling; it demonstrates a different approach to establishing responsibility before punishment.

The strongest lesson is institutional as well as technical. Courts accepting adversarial documents into an AI workflow need to retain their status as untrusted source material, preserve the original evidence and keep a human accountable for decisions. Our information-flow-control lesson explains the related problem of stopping source content from acquiring authority over later actions.

The ruling is not verified as the world’s first, and November circulation should not turn May’s decision into a new judgment. Its significance survives those limits: concealed model-directed instructions have already reached a real court, been detected in an AI-assisted workflow and produced a substantial sanction without proven success of the attack.

Primary source, verified: read the paper →

Ground Truth. (2026, November 1). A Brazilian judge fined two lawyers over a hidden AI instruction in a court filing. Ground Truth. https://groundtruth.day/news/brazil-hidden-court-prompt-injection-sanction.html

@misc{groundtruth:brazil-hidden-court-prompt-injection-sanction,
  title  = {A Brazilian judge fined two lawyers over a hidden AI instruction in a court filing},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/brazil-hidden-court-prompt-injection-sanction.html}
}

Topics: cybersecurity · prompt-injection · ai-security · courts · governance