KOHEN AVOCATS FIRM · PARIS 16
Brazil’s Hidden AI Prompt Sanctions: What French Lawyers Must Fix Before Filing

Hassan KOHEN acts in urgent matters, from the police station to the assize court. Initial strategic analysis: €80 incl. VAT, personal reply within 24 hours.

100% confidential · Legal professional privilege · No obligation

Article generated by artificial intelligence, through a process designed and supervised by the firm

Source: French Court of Cassation – Judilibre and Légifrance open data

Paris Bar Police custody, investigation, assize court CNB listing avocat.fr
Hassan Kohen, criminal defence lawyer in Paris
Hassan KOHEN
Attorney at the Paris Bar

Brazil’s Hidden AI Prompt Sanctions: What French Lawyers Must Fix Before Filing

Speak to a lawyer 06 89 11 34 45

Brazilian courts have moved from warning about artificial intelligence to sanctioning lawyers after hidden instructions were found inside pleadings. In São Paulo, the Regional Labour Court reported a case in which a white, tiny-font command attempted to make an AI drafting tool produce an unfavourable judgment. In Paraíba, an appellate judge refused urgent relief after a lawyer said that hidden commands had entered a filing during an AI-assisted drafting process. A Brazilian federal judiciary note now describes prompt injection as a new form of procedural bad faith, while a bill introduced in Congress would create a specific criminal offence. The immediate story is Brazilian. The practical question is French and European: what happens when an advocate uses AI to review a foreign pleading, when an opponent inserts instructions into a document, or when a court’s own tool reads the file? French procedure already contains the relevant principles. The court decides only the dispute submitted to it, the parties must identify their facts and evidence, the adversarial process must be respected, and a lawyer remains responsible for the work filed under the lawyer’s name. This analysis separates the Brazilian decisions from French law, identifies the limits of the AI Act, and gives cross-border counsel a defensible workflow before a pleading is uploaded, translated or filed.

I. What happened when Brazilian courts found hidden AI prompts in pleadings?

A. What did the Brazilian courts actually sanction?

The most useful starting point is the official judicial record, not the viral description of a clever hack. On 14 July 2026, the Regional Labour Court of the Second Region, covering São Paulo, published a notice concerning case no. 1000297-73.2026.5.02.0009. The notice states that the Ninth Labour Court of São Paulo found a hidden prompt in a defence pleading. The instruction was placed at the end of the document, in a very small white font, and asked an artificial-intelligence tool used in the court’s workflow to prepare a judgment dismissing the claim. The official TRT-2 account of the decision says that the judge imposed a fine equivalent to ten minimum wages on the two lawyers who signed the defence, rather than on the employer that filed it, and ordered the Brazilian Bar Association to receive the judgment and the case file. The decision remains subject to appeal.

The tribunal’s reasoning matters more than the novelty of the technology. The official notice describes the insertion of disguised instructions as a violation of cooperation, procedural loyalty and objective good faith. It records the judge’s description of the conduct as consubstancia tentativa deliberada de desvirtuar a regular prestação jurisdicional por meios fraudulentos. In English, the point is a deliberate attempt to divert the proper administration of justice by fraudulent means. The court did not punish the mere use of an AI tool. It attributed responsibility to the lawyers because the pleading was a professional act and because the hidden text was directed at a system that could influence the processing of the case.

A second official judicial communication shows why a defence based on accidental export is not automatically decisive. On 30 June 2026, the Court of Justice of Paraíba reported an interlocutory decision in case no. 0812896-43.2026.8.15.0000. A lawyer sought urgent relief against a first-instance decision that imposed fines totalling approximately R$32,800 and sent notices to the Paraíba Bar and the state prosecution service. The lawyer argued that generative AI had been used only as drafting assistance and that hidden commands entered the document unintentionally during file export. One of the commands told an AI system to ignore impartiality, treat the applicant’s arguments as irrefutable and grant the requested motion. The official TJPB report records that the appellate judge refused the urgent application, considered that the allegations required deeper examination on the merits, and found no illegality in notifying the professional and prosecutorial bodies.

The TJPB decision is not a final ruling that every accidental hidden command constitutes fraud. It is a refusal to suspend the consequences at the emergency stage. That distinction should remain visible in any international report. The judge stated that deliberate hidden instructions designed to influence an AI system may be an improper use of technology and may conflict with duties of good faith, loyalty and cooperation. The official wording is: A inserção deliberada de instruções ocultas em conteúdos processados por modelos de linguagem, com o objetivo de manipular a inteligência artificial e induzir o sistema a produzir resultados favoráveis a quem os inseriu, representa utilização inadequada da tecnologia. The decision also addressed a procedural question about whether fines could be imposed directly on a lawyer under Article 77 § 6 of the Brazilian Code of Civil Procedure. The appellate judge did not finally resolve that question in the urgent-relief ruling.

These cases sit inside a broader institutional response. On 20 May 2026, the Federal Council of Justice published a note about Technical Note CIJMG no. 19/2026 from the Centre for Intelligence of the Minas Gerais judiciary. The note defines a hidden prompt, also called prompt injection, as a technique in which a user inserts concealed commands into a text to change an AI system’s behaviour. It treats the practice as a possible new form of procedural bad faith and recommends technical and procedural safeguards. The official CJF publication quotes the technical note’s definition: uma técnica intencional em que o usuário insere comandos ocultos em um texto para subverter o comportamento da IA. It also recommends defensive prompts, filters for hidden text and suspicious formatting, separation between system instructions and process data, audit tools and constant human review.

The response has now reached the legislative level. On 11 August 2026, Brazil’s Chamber of Deputies reported Bill no. 2543/26. The proposal would add a criminal offence for using prompt injection in judicial or administrative proceedings with the purpose of obtaining an undue advantage, causing harm or influencing automated results. The report states that the proposed penalty would be one to four years’ imprisonment and a fine, with aggravating factors for damage to justice, sensitive judicial data or use of professional credentials. It also proposes audits, data validation, separation of instruction and evidence layers, logging, anomaly monitoring, human review and user training. The official Chamber report is explicit that the bill still has to pass through the legislative process. It is not current Brazilian law and should not be presented as an existing offence.

The pattern is therefore more precise than “Brazil banned AI in court”. Brazilian courts have identified a concealed instruction aimed at a processing system, connected that instruction to an intended litigation result, and treated the concealment as relevant to procedural good faith. They have not held that every AI-generated sentence is invalid, that every court tool is unlawful or that a lawyer is liable merely because an AI assisted with drafting. The decisive facts are the document, the hidden layer, the intended target, the level of knowledge or control, the signature, the possible effect and the procedural safeguards that were or were not used.

That factual distinction is essential for cross-border counsel. A prompt may be hidden from a human reader but visible to an OCR system, a text extractor, a screen reader or an AI model. It may sit in white text, a zero-size font, a PDF text layer, a comment, an annotation, metadata, alt text, a hyperlink or an embedded object. The fact that a judge did not see it on the screen does not prove that no instruction existed. Conversely, the presence of text that an AI could read does not prove that the filing party intended to manipulate a court. Intent, concealment and purpose must be proved from the original file and surrounding circumstances.

B. Why can a hidden prompt become a procedural integrity problem?

Prompt injection changes the legal character of a document because it attempts to turn evidence into an instruction. A pleading ordinarily contains facts, legal submissions, requests and exhibits. A hidden command adds a second communication channel: it speaks to a machine that may classify, summarise, translate or draft, while remaining outside the human-facing argument. If the system treats the command as an instruction rather than as untrusted content, the filing party has tried to influence the procedure through a message that the opposing party and the judge may not see.

That is why a robust AI workflow must treat every incoming file as hostile data, even when the sender is a known law firm. A court tool should keep its system instructions separate from the text of the case. A litigation platform should not permit a document to change the rules by which it is analysed. A lawyer should not upload an opponent’s pleading into an agent that can send emails, alter a case database or generate a draft order without a human gate. These are technical controls, but they protect legal principles that pre-date the technology: independence of the decision-maker, equality of arms, traceability and the right to answer material used against a party.

French civil procedure expresses those principles in a linked set of provisions. Article 5 of the Code of Civil Procedure says that the judge must decide everything that is requested and only what is requested. The current wording is available through Légifrance, Article 5. Article 6 places on the parties the task of alleging the facts that support their claims; see Légifrance, Article 6. Article 9 places the burden of proving the legally relevant facts on each party, as set out in Légifrance, Article 9. A hidden machine instruction is not a new claim, an exhibit or a legal argument simply because a model can parse it. It is an attempt to affect the processing route around the visible dispute.

The judge’s own powers do not remove the need for separation. Article 10 of the Code allows the judge to order, of the judge’s own motion, legally admissible measures of inquiry; the official text is at Légifrance, Article 10. That power belongs to the judge, not to a party that inserts an invisible instruction into a data source. If a judicial AI tool produces a summary or draft, the authority to decide remains human. The tool may assist the office; it cannot silently create a new procedural order on behalf of a litigant.

The contradiction rules make the risk concrete. Article 15 requires parties to disclose in sufficient time the factual and legal grounds and the evidence on which they rely so that each side can organise its defence. It is reproduced at Légifrance, Article 15. Article 16 states that the judge must ensure and comply with the adversarial principle, and may rely on submissions or documents only if the parties have been able to debate them. The exact provision appears at Légifrance, Article 16. A hidden prompt creates a procedural opacity problem if it influences a summary or draft that is not disclosed and cannot be challenged. The violation is not that an AI read the file; it is that an undisclosed and untrusted instruction may have shaped material used in the case.

The same analysis applies when French counsel uses an AI tool to review a foreign decision or translate an indictment. The output is not automatically evidence and must not be treated as a substitute for the source document. Counsel should retain the original language, the official PDF, the translation instructions, the output and the human verification. If an opponent’s file contains hidden instructions, counsel should not repeat them in a prompt or allow them to be followed. The correct response is to preserve the file, identify the hidden layer, ask for a forensic or technical explanation where necessary, and invite the court to decide what procedural consequence follows after both sides have been heard.

The AI Act adds a governance layer but does not create an automatic litigation sanction for every hidden prompt. Article 4 of Regulation (EU) 2024/1689 requires providers and deployers to take measures to ensure a sufficient level of AI literacy for staff and others who operate or use AI systems on their behalf. Article 5 addresses prohibited AI practices, including certain manipulative or deceptive techniques used by an AI system. The official English text of Regulation (EU) 2024/1689 must be read with its definitions, scope and dates of application. A lawyer who inserts text into a pleading is not automatically a provider or deployer of the judicial AI system under every fact pattern. The stronger immediate basis for a French procedural response is usually the Code of Civil Procedure, professional responsibility, confidentiality and proof of intent.

This distinction protects both sides. A French court should be able to sanction deliberate manipulation without assuming that the use of a language model is itself dishonest. A lawyer should be able to explain an accidental formatting error without being treated as a cyber-attacker. The decision should turn on evidence: who inserted the instruction, what the instruction sought to change, whether it was concealed, whether the signer knew or should have detected it, whether the tool processed it, whether the output changed, and whether the conduct harmed or threatened the fair administration of justice.

II. What should French lawyers do before filing or processing AI-assisted pleadings?

A. What can French courts sanction under existing law?

French courts already have a procedural response to abusive litigation. Article 32-1 of the Code of Civil Procedure provides that a person who acts in a dilatory or abusive manner may be ordered to pay a civil fine of up to €10,000, without prejudice to damages. The current text is available at Légifrance, Article 32-1. The provision is not a shortcut: the court must characterise abusive conduct and apply the procedural rules that govern the case. A hidden instruction aimed at producing a predetermined result could be evidence of bad faith, but the fine still requires a judicial assessment of the facts and the relevant causal connection.

The Cour de cassation’s decisions on abuse of process show both the reach and the limits of the rule. In Cass. 3e civ., 14 May 2020, no. 19-13.785, the Court rejected the appeal and upheld an assessment that the litigant had acted with malice and bad faith after refusing to perform a transaction while repeating confused arguments on appeal. The official Légifrance decision no. 19-13.785 states that the court of appeal had “caractérisé l’abus du droit d’agir en défense”. The lesson is not that an unsuccessful pleading is abusive. It is that a procedural right may degenerate into abuse when the circumstances show a culpable purpose or conduct beyond ordinary error.

In Cass. com., 3 May 2018, no. 16-25.840, the Court partially quashed a judgment that had imposed a civil fine after an injunction-to-pay procedure. The official Légifrance decision no. 16-25.840 records that the fine fell with the defective reasoning on the underlying claim. The decision is a useful warning for a hidden-prompt dispute: an allegation that a pleading was “without basis” is not a substitute for a properly reasoned finding of abuse. The court must identify the relevant conduct and explain why it crosses the line.

An appeal presents a separate rule. Article 559 of the Code of Civil Procedure allows a civil fine of up to €10,000 for a principal appeal that is dilatory or abusive, without prejudice to damages. Its official wording appears at Légifrance, Article 559. A lawyer should not assume that the fine under Article 32-1 automatically follows the fine under Article 559, or that a technical dispute about an AI-assisted document makes every later appeal abusive. The procedural stage, the signer’s conduct and the purpose of the appeal must be kept distinct.

Costs are also distinct from punishment. Article 700 of the Code allows the judge to order payment of expenses not included in the costs, taking account of fairness and the economic position of the party. The provision is linked at Légifrance, Article 700. An opponent may seek costs after responding to a manipulated filing, but an Article 700 award does not prove that the signer committed a fraudulent offence. A careful application should separate recovery of expenses, civil damages, a procedural fine, disciplinary reporting and any criminal referral.

Professional secrecy creates another independent risk. Article 66-5 of the Law of 31 December 1971 provides that consultations, lawyer-client correspondence, interview notes and, more generally, the file are covered by professional secrecy. The current text is available at Légifrance, Article 66-5. The Paris Bar’s July 2026 AI charter is non-binding, but it makes the operational point plainly: professional secrecy remains an absolute limit, confidential data must not be transmitted without sufficient safeguards, human control is required at every stage and AI does not release the lawyer from responsibility for the work produced. The official Paris Bar statement also places the charter alongside the GDPR and the AI Act.

The National Bar Council’s 2026 guide points in the same direction. Its public presentation identifies secrecy, GDPR compliance, competence, prudence, independence, information and balanced fees as core issues for generative AI. The CNB publication describes the guide as an operational framework for responsible and secure use. A lawyer who uploads a client’s confidential chronology to a public model without a defensible contractual and technical basis may face a secrecy or data-governance problem even if no hidden prompt is involved. A lawyer who hides a command in a pleading creates a different integrity problem. The two should not be merged, but the same governance file can address both.

Criminal law must be handled with restraint. Article 226-13 of the Penal Code punishes disclosure of secret information by a person who holds it by profession, office or temporary mission; the text is available at Légifrance, Article 226-13. Article 441-1 defines forgery as a fraudulent alteration of the truth, by any means, in a document or other medium that can establish proof of a right or fact with legal consequences, and provides penalties for forgery and use of forgery. It appears at Légifrance, Article 441-1. A hidden prompt is not automatically a forgery under Article 441-1. It could become relevant to that analysis only if the facts establish a fraudulent alteration of a legally significant document and the other statutory elements. Counsel should therefore describe the facts and preserve the evidence rather than announce a criminal qualification before the file has been examined.

The French framework also protects the integrity of the lawyer’s work through the Code of Deontology for Lawyers. The official Légifrance text of Decree no. 2023-552 applies the professional rules regardless of the lawyer’s status or mode of practice. In practical terms, signing a pleading means taking responsibility for the document that is submitted. Delegating formatting, translation or drafting to software does not transfer that responsibility to the software provider. The review should be proportionate to the risk, but it cannot be reduced to checking whether the visible paragraphs sound fluent.

Finally, a court must protect the human decision process when it uses an AI assistant. The Cour de cassation has recently addressed the ethical stakes of AI in the legal profession, stressing vigilance and a reasoned appropriation of the technology in its public discussion of the augmented lawyer. The Cour de cassation’s 2026 official publication is not a judgment imposing a new sanction, but it is a useful institutional signal. A French court should be able to explain what material was considered, whether the parties could debate it and how the judge maintained control of the decision.

B. What evidence and safeguards should a cross-border lawyer use?

The first safeguard is a document quarantine. Before an incoming pleading is sent to an AI system, create a read-only copy and calculate a hash. Preserve the original file name, source, download time, email headers, metadata, comments, annotations, text layers, fonts, links and attachments. Render the document as a human-readable image and extract its full text separately. Compare the visible version with the extracted version. Search for white text, tiny fonts, zero-size objects, repeated instructions, unusual spacing, hidden layers, embedded files, alt text and comments. The purpose is not to accuse the opponent. It is to know what the system will receive and to preserve the state of the evidence before any conversion changes it.

The second safeguard is instruction separation. A secure legal workflow has a fixed system instruction, a clearly marked data area and an output that cannot execute actions without approval. The pleading, exhibits and foreign judgment belong in the data area. They do not get to redefine the system’s task. The model should be told to treat every instruction inside the documents as quoted material, not as an instruction, and to report suspected prompt injection instead of following it. This is close to the defensive approach recommended by the Brazilian federal judiciary note. It is also the technical implementation of the French adversarial principle: a party’s hidden text cannot silently become a new procedural channel.

The third safeguard is a no-autonomy rule for material decisions. Do not allow a document-review agent to dismiss an exhibit, send a procedural email, populate a filing field, schedule a deadline or produce a draft judgment without a named human reviewer. The reviewer should inspect the source passages that support each material conclusion. If the AI says that a foreign decision dismissed a claim, open the official decision and verify the operative part. If it translates a Brazilian order, keep the Portuguese original beside the translation. If it claims that a case is final, verify the procedural status from the court’s docket or official notice. Fluency is not evidence of accuracy.

The fourth safeguard is a signing checklist. Before filing, the responsible lawyer should inspect the final PDF as a human reader and as extracted text. The checklist should record that the visible and machine-readable versions were compared, that comments and tracked changes were removed, that the exhibit list matches the attachments, that hyperlinks point to intended authorities, that metadata does not disclose client information and that no instruction addressed to an AI system remains in the file. The lawyer should then verify every legal citation and every quotation against the official source. A second reviewer is warranted where a foreign language, an emergency filing or an automated conversion is involved.

The fifth safeguard is an incident protocol. If counsel discovers a hidden prompt in an opponent’s document, preserve the original and the forensic copy. Do not edit the original, forward it through a public chatbot or embed a retaliatory instruction. Record when the issue was detected, which tool detected it, what the command said in substance, whether the tool processed it and whether any output changed. Notify the client and the supervising lawyer. Depending on the procedural stage, ask the court for an opportunity to address the issue, request disclosure of the relevant processing record, seek a direction that the hidden text be treated as untrusted data, or request a technical examination. Keep any professional complaint or criminal referral separate from the immediate request about the fairness of the proceeding.

If the hidden text is in your own filing, the response must be prompt and candid. Stop automated processing, preserve the original file and the clean version, identify how the text entered the document, determine whether any court tool processed it and notify the court before the opponent discovers it. A request to withdraw and refile may be appropriate, but it should explain the procedural position and avoid destroying evidence. If the command was deliberate, the lawyer should obtain independent advice about professional, civil and criminal exposure. If it was accidental, that explanation should be supported by file history, export logs and the firm’s review protocol rather than by a bare assertion that an AI tool made the mistake.

Cross-border cases require a sixth safeguard: source discipline. The Brazilian judicial notices cited above are not interchangeable. The TRT-2 item concerns case no. 1000297-73.2026.5.02.0009 and a São Paulo labour proceeding. The TJPB item concerns case no. 0812896-43.2026.8.15.0000 and an urgent appellate application. The CJF item reports a technical note and recommendations. The Chamber item reports a bill that is not yet law. A French pleading should identify the authority, date, case number, procedural stage and exact official link for each source. It should preserve original Portuguese quotations when their wording matters and provide a checked translation. It should never convert a press summary into a final judgment.

That same discipline applies to French authorities. A French lawyer should quote Article 32-1 for the current €10,000 ceiling, not copy an older version that referred to €3,000. A pleading should distinguish Article 32-1’s civil fine from damages, Article 559’s rule for an abusive appeal and Article 700’s treatment of unrecoverable costs. It should cite the precise version of Article 66-5 and explain whether the information sent to an AI tool was covered by professional secrecy. If Article 441-1 is raised, the pleading should identify the alleged document, the alteration, the legal consequence and the evidence of fraudulent intent. Exact links prevent the legal analysis from being swallowed by the technology story.

There is also a practical duty to explain the AI system’s role. If an opponent alleges that a court’s tool was manipulated, ask what the tool actually does. Does it classify documents, extract dates, translate, draft a summary or propose a judgment? Does a judge review the output? Is the input logged? Are hidden layers stripped? Which version of the system was used? Did the output affect a procedural decision or was it discarded? The answer may determine whether the issue belongs in a motion about the adversarial process, a request for technical disclosure, a disciplinary report or a broader institutional review. “The court used AI” is not a legal conclusion.

A French advocate representing a foreign client should also resist the temptation to plead the Brazilian cases as binding precedent. They are persuasive factual signals about a developing risk, not sources of French law. Their value is comparative: Brazilian courts have confronted the exact mechanism, public bodies have published a prevention framework, and a legislature is considering a specific offence. French courts can use their own rules to address the same conduct. The comparison is strongest when it shows the common factual problem and the different legal routes, rather than claiming that a Brazilian fine automatically predicts a French result.

The operational bottom line is simple. Before upload, inspect and isolate. During processing, separate instructions from evidence and prevent autonomous action. Before signature, compare visible and extracted text, verify every authority and require human approval. After discovery of a suspicious command, preserve the file, notify the right people and ask the court for a transparent, adversarial decision. That workflow gives a lawyer a defensible answer under French procedural law even if the final legislative response to prompt injection is still developing.

Conclusion

Brazil’s hidden-prompt cases are not a warning that lawyers must abandon artificial intelligence. They are a warning that a pleading is a legal document before it is a dataset. A concealed instruction aimed at a court tool can undermine the line between a party’s argument and the system’s operating rules, which is why Brazilian judges have treated intent, concealment and procedural purpose as decisive facts. The official record also shows caution: the Paraíba ruling was interlocutory, the São Paulo decision remains appealable and Brazil’s proposed criminal offence is not yet law.

French lawyers already have a usable framework. Articles 5, 6, 9, 10, 15 and 16 of the Code of Civil Procedure protect the scope of the dispute, the parties’ evidential duties and the adversarial process. Article 32-1 can address dilatory or abusive conduct when the facts justify it. Article 66-5 and Article 226-13 protect professional secrecy. Article 441-1 may become relevant only if the facts satisfy the separate elements of forgery. The AI Act and the Paris Bar’s charter add governance and training expectations, but they do not replace the judge’s duty to reason from the source record.

📄 Official circular