STJ Opens Police and Administrative Probe Into 'Prompt Injection' Attacks on Court AI

Overview

• On May 20 the Superior Tribunal de Justiça ordered a police inquiry and an internal administrative procedure after technicians flagged documents with hidden AI commands in cases under the court's review.
• Investigators are reviewing at least 11 flagged criminal cases from multiple states and plan to take depositions of implicated lawyers and firms to determine whether to apply procedural, administrative or criminal penalties.
• Prompt injection is a technique that hides instructions inside filings—for example white text on a white background—that tries to force an AI assistant to produce shallow or biased analysis rather than conduct a full legal review.
• The STJ said its generative-AI tool STJ Logos and the TRT-4's assistant Galileu include three layered defenses—preprocessing to isolate malicious input, context-scope limits, and output compliance filters—that neutralized the detected injections.
• The probes follow journalistic inventories and earlier incidents such as the Parauapebas labour case and reports that one lawyer filed nine documents with identical hidden commands; the developments are pushing courts to certify injection attempts in case files and strengthen audits, prompt banks and algorithmic training for legal staff.