The Rise of Prompt Injection in Brazil's Judicial AI Systems
Brazil's judiciary, one of the busiest in the world with over 80 million pending cases, has embraced artificial intelligence tools to manage workloads. Systems designed for case triage, document summarization, and decision drafting are now standard in many courts. Yet recent incidents have highlighted a sophisticated threat: invisible manipulation through prompt injection, where hidden instructions embedded in court filings attempt to steer AI outputs without detection by human readers.
This technique involves placing text in white font on white backgrounds, microscopic font sizes, or other invisible formats within digital documents. When AI models process these files, they interpret the concealed commands as legitimate instructions, potentially biasing analysis or recommendations. Lawyers in at least one regional labor court have faced sanctions after such attempts came to light, underscoring the judiciary's vulnerability.
The National Council of Justice, known as CNJ, has responded with guidelines under Resolution No. 615/2025. These rules emphasize transparency, human oversight, and security measures for AI deployment across Brazil's court system. The National Committee on Artificial Intelligence in the Judiciary, or CNIAJ, has issued technical notes addressing adversarial attacks like prompt injection, calling for robust defenses.
Documented Cases of Hidden AI Interference
One prominent example emerged in the Regional Labor Court of the 8th Region, or TRT-8. Filings contained concealed prompts designed to influence the court's AI-assisted review processes. Court officials detected the manipulation, leading to professional repercussions for those involved. Similar tactics have been reported in other jurisdictions, including attempts to alter how AI evaluates evidence or prioritizes arguments.
In the Supreme Federal Court, or STF, Claim No. 78,890 revealed fabricated precedents in a petition. While not explicitly tied to invisible text, the case illustrated broader risks of AI-generated or AI-influenced content entering proceedings. Justice Flávio Dino rejected the claim and imposed sanctions for bad-faith litigation, highlighting the need for verification mechanisms.
Another incident involved a Santa Catarina court where a lawyer admitted using generative AI to create nonexistent case law in an appeal. The Court of Justice of Santa Catarina imposed a 10 percent fine for bad-faith litigation, reinforcing that courts will not tolerate misleading AI-assisted submissions.
Broader Context of AI Adoption in Brazilian Courts
The Brazilian judiciary's push toward AI stems from chronic backlogs. Tools such as Victor at the STF, JuLIA, and various regional systems assist with initial screening, tax enforcement, and jurisprudence comparison. Generative AI applications, including the STF's MARIA system for document drafting, have accelerated routine tasks while judges retain final authority.
Reports from the Center for Justice at Fundação Getulio Vargas, or FGV Justiça, document 178 active AI projects across 58 courts as of recent surveys. Generative AI appears in nearly half of tribunals, with many more planning adoption. Productivity gains in tax cases have reached up to 67 percent in some instances, yet experts stress that these systems remain assistive rather than autonomous.
Challenges include algorithmic bias, hallucinations where AI invents facts or citations, and opacity in decision-making processes. The Portulans Institute case study on AI in Brazilian courts notes that undisclosed generative AI use in drafting has already prompted nullity requests and evidentiary challenges.
Photo by Kseniia Lobko on Unsplash
Stakeholder Perspectives on the Emerging Risks
Judges and court administrators express cautious optimism tempered by vigilance. Many welcome efficiency improvements but warn against over-reliance. The CNJ stresses that final decisions must always rest with human magistrates, with AI outputs subject to rigorous review.
Legal practitioners highlight the ethical tightrope. While generative tools can speed research, the temptation to insert subtle manipulations or rely on unverified outputs raises professional responsibility questions. Bar associations have been notified in several cases involving fabricated content.
Technology experts and governance bodies point to the need for adversarial testing. The CNIAJ's proposed Program for Adversarial Security for Artificial Intelligence Systems in the Brazilian Judiciary aims to simulate attacks and harden defenses. Public trust in the justice system depends on demonstrable safeguards against both intentional manipulation and unintentional errors.
Regulatory and Technical Responses Underway
Beyond Resolution 615/2025, Brazilian authorities have advanced complementary measures. Electoral courts have banned AI chatbots from offering voting advice and shifted the burden of proof in disputes involving manipulated content. Platforms must disclose AI-generated material during election periods.
Technical solutions under discussion include watermarking AI outputs, enhanced document parsing that detects hidden text, and mandatory audit trails for AI-assisted filings. Some courts now require parties to declare any AI use in submissions.
International observers, including the International Bar Association, have examined Brazil's experience as a leading case study in AI-generated evidence and manipulation risks. Lessons from these incidents inform global conversations on judicial technology governance.
Implications for Fairness, Transparency, and Access to Justice
Invisible manipulation threatens core principles of due process. If AI recommendations can be covertly influenced, parties with greater technical resources gain unfair advantages. This dynamic could exacerbate inequalities in an already complex legal environment.
Transparency requirements aim to mitigate these risks. Public reporting on AI tool performance, bias audits, and incident disclosures help maintain accountability. Continued investment in judicial training ensures that magistrates can critically evaluate AI suggestions.
Longer term, unchecked vulnerabilities could erode public confidence in court outcomes. Brazil's experience demonstrates that technological adoption must proceed hand in hand with ethical frameworks and technical resilience.
Photo by Kseniia Lobko on Unsplash
Future Outlook and Recommended Safeguards
As AI capabilities evolve, Brazilian courts are positioned to lead in responsible implementation. Expanded use of the SINAPSES platform and Justice 4.0 Program offers opportunities for standardized, secure deployment across jurisdictions.
Recommended safeguards include regular penetration testing of AI systems, clear protocols for disclosing AI assistance, and interdisciplinary collaboration between legal experts, data scientists, and ethicists. International partnerships can accelerate the sharing of best practices.
Stakeholders agree that the goal remains enhancing access to justice without compromising integrity. With proactive measures from the CNJ and CNIAJ, Brazil can transform potential threats into opportunities for more robust, equitable judicial processes.
Practical Steps for Legal Professionals and Court Users
Anyone interacting with Brazilian courts should verify the authenticity of all submissions and avoid any form of hidden formatting. Legal teams are advised to document AI tool usage and cross-check all generated content against primary sources.
Court users can consult official CNJ resources for guidance on AI policies. Staying informed about evolving regulations helps prevent inadvertent violations and supports the integrity of proceedings.
Ultimately, awareness of these invisible risks empowers better practices. The judiciary's commitment to transparency and oversight provides a foundation for continued innovation that serves all Brazilians fairly.