On May 19, the CIJMG issued Technical Note N. 19/2026, warning judges about an emerging form of procedural fraud: the intentional manipulation of AI tools through “prompt injection”.
This involves intentional misconduct in which hidden commands are inserted into pleadings or procedural documents to influence the output generated by the Judiciary’s AI systems, whether in case summaries or draft decisions.
The Technical Note describes concrete methods of insertion: invisible text, zero-width characters, instructions embedded in PDF and Word metadata, or commands placed at the end of lengthy documents. Content that is imperceptible to the human reader, but still processable by language models.
From a legal perspective, such conduct may constitute bad-faith litigation (art. 80 of the Brazilian Code of Civil Procedure), an act against the dignity of justice (art. 77 of the Brazilian Code of Civil Procedure), and, in more serious cases, procedural fraud (art. 347 of the Brazilian Criminal Code).
As an institutional response, the CIJMG recommends the adoption of defensive prompts by judges, mandatory human oversight over AI-generated content (the so-called human-in-the-loop protocol), and technical measures such as document sanitization filters and protection of the system prompts used by the Court’s AI tools.
The full text is available at the following link: CIJMG Technical Note No. 19/2026.
