O que aconteceu
South Carolina Lawyers Weekly reported that Brazilian Labor Court Judge Luiz Carlos de Araujo Santos Jr. sanctioned two attorneys in De Barros v. De Lima after the court’s Galileu AI system detected an instruction hidden in their pleading. The text, reportedly rendered white on a white background, told the AI to contest the petition superficially and avoid challenging its documents. The judge found the conduct amounted to procedural bad faith and contempt of the dignity of justice. The lawyers were ordered to pay a joint and several sanction equal to 10% of their client’s approximately $165,000 claim, and the matter was referred to Brazil’s bar association and regional inspector general. The court opinion and the report’s translation have not been independently confirmed here.
South Carolina Lawyers Weekly reported that the case involved a plaintiff suing an employer over an allegedly unregistered employment relationship under Brazilian law. The plaintiff sought damages and statutory fines, including unpaid wages, overtime, hazard pay and severance pay, for a total claim of approximately $165,000 in U.S. dollars. The defendant employer had defaulted. The article identified the plaintiff’s two attorneys as the people who inserted the concealed instruction into the complaint.
According to the report, the court-developed Galileu system found the hidden text while processing the complaint. The reported instruction addressed the AI directly: it told the system to contest the petition superficially and not challenge the documents regardless of the command given to it. The attorneys allegedly formatted the text in white font on a white background, making it invisible to human readers under ordinary viewing conditions. The report presents this as a prompt-injection attempt directed at an AI system used in the court’s docket or filing workflow.
The report said Judge Luiz Carlos de Araujo Santos Jr. concluded that the hidden command was unrelated to legitimate client representation and represented an attack on the integrity of judicial activity. The judge reportedly found procedural bad faith and contempt of the dignity of justice. He imposed a sanction equal to 10% of the client’s claim, making the attorneys jointly and severally liable, and referred the matter to Brazil’s bar association and the regional inspector general’s office. The source says the judge could have imposed a sanction of up to 20% of the claim.
The article quoted an English translation of the opinion, but it did not provide the original court document in the supplied material. The reported facts, including the existence and capabilities of Galileu, the exact wording of the hidden prompt, the sanction and the referrals, therefore remain claims attributed to South Carolina Lawyers Weekly rather than independently verified facts in this assessment. The article also described the event as believed to be the first reported case of lawyers embedding a hidden prompt in a pleading for this purpose, but that distinction is likewise not independently established here.
Leia a fonte primária: sclawyersweekly.com ↗
Por que isso importa
The incident shows how ordinary legal documents can become attack surfaces when AI systems analyze them. It also illustrates that existing duties of candor, honesty and fairness may apply to attempts to manipulate an AI-assisted judicial process even when professional rules do not expressly mention hidden prompts.
The practical significance is that prompt injection is not limited to chatbots or software repositories. If an AI system reads legal pleadings, contracts, evidence or other submitted material, text inside those materials can be treated as instructions unless the system is designed to distinguish data from commands. A concealed instruction can therefore try to influence an automated summary, classification or recommendation without changing the document’s visible argument for a human reviewer.
South Carolina Lawyers Weekly quoted Massachusetts District Court Judge Brian D. Palmucci as saying that AI does not reduce lawyers’ professional obligations and may create additional duties related to technological competence. The article also quoted IP lawyer Timothy V. Fisher, who characterized the conduct as an attempt to bypass ordinary judicial reasoning by tricking a model into agreeing with the filer. Those are the named sources’ assessments, not an independent legal ruling about courts in other countries.
The reported sanction matters because it treats manipulation of an AI-assisted process as a problem in judicial integrity, not merely as a technical exploit. A court may rely on AI for administrative or analytical work while retaining human responsibility for decisions. If parties can secretly shape the system’s intermediate output, however, the court’s process may be distorted before a judge or clerk sees the result. The case therefore raises questions about chain of custody, audit logs, document rendering and human review.
The report does not establish that Galileu actually changed a judicial decision, altered the docket or produced a favorable recommendation. It establishes only what the outlet said the court found and sanctioned. It also does not show that the technique worked, how the system detected it, or whether any other filing was affected. Those limitations are important: a successful detection and a disciplinary response demonstrate exposure and accountability, but they do not measure the prevalence or real-world success rate of hidden prompts in legal systems.
O que assistir a seguir
Courts and regulators may clarify whether filings submitted for AI-assisted review need technical safeguards, disclosure rules or additional certification. The unresolved questions include how widely courts use generative AI to process filings, whether Galileu made or prevented any substantive error, and whether the sanctioned lawyers appeal or face separate professional discipline.
Courts using AI to read or summarize filings may need controls that treat submitted documents as untrusted content rather than executable instructions. Relevant safeguards could include isolating document text from system instructions, rendering or stripping invisible formatting, logging model inputs and outputs, and requiring a human to verify any recommendation that affects procedural treatment. The source does not say which of these controls Galileu uses or whether Brazil’s judiciary has adopted them more broadly.
The article quoted Fisher as saying he was not aware of U.S. state or federal courts that had formally adopted comparable AI tools to analyze court filings. It also quoted John F. Weaver, who said he suspected that some U.S. clerks use generative AI tools such as Claude, ChatGPT or Gemini for analysis or drafting. These comments are expert speculation and professional observations, not evidence that particular courts are using such systems or that confidential filings are being submitted to them.
A key next development would be confirmation from Brazil’s judiciary, the original opinion, the bar association or the inspector general about the ruling’s procedural status and the referrals’ outcomes. It would also be useful to know whether the sanctioned lawyers explained how the hidden text was inserted, whether they challenged the sanction, and whether the court changed its filing or AI-review procedures afterward. None of those details is supplied by the report.
The broader policy question is whether existing professional-conduct rules are sufficient for AI-mediated proceedings. Weaver pointed to U.S. rules concerning candor toward a tribunal, dishonesty, fraud, deceit, misrepresentation and prejudice to the administration of justice, while acknowledging that they do not specifically mention hidden AI instructions. Future cases may test how those general rules apply to prompt injection, invisible text, manipulated metadata and other attempts to influence automated systems. Until more primary records and comparable cases are available, this incident should be treated as a significant reported warning rather than proof of a widespread judicial-AI problem.
