A judicial filing typically conveys an argument, evidence, or a request for relief. However, once a court uses artificial intelligence to process a filing, the same document may also interact with the computational intermediary through which judicial work proceeds. A 2026 Brazilian labor case illustrates this distinction: A pleading contained visually concealed language directed to the court’s AI system. The system detected the instruction, alerted the judge, and counsel was sanctioned. The public record does not indicate whether the instruction was followed or if the judgment changed.
This incident reveals a gap between procedural law, which governs participants and submissions, and AI regulation, which governs system design, deployment, and security. Prompt injection through an authorized filing channel engages both bodies of law because the submitter uses a lawful procedural role to seek computational authority that the institution never conferred.
The article responds in three steps. First, it develops a legal taxonomy of adversarial filings and related inputs. Second, it compares the Brazilian incident under the European Union Artificial Intelligence Act and United States law. Finally, it proposes a Two-Sided Adversarial Filings Framework. This framework combines participant-side responsibilities for unauthorized machine-directed content with institutional responsibilities for reasonable adversarial preparedness. The framework calibrates sanctions and procedural remedies according to the demonstrated path from exposure to compromise, procedural influence, decisional effect, and systemic propagation. Therefore, participant culpability neither excuses institutional insecurity nor depends upon a successful attack.