The court receives a filing as evidence and argument, not as instructions for its own software. In the reported petition, text rendered effectively unreadable to people was still present in the document extraction layer, with imperative language asking an automated reader to accept the appeal. That is an attempted change of processing authority through…
Inspect the ClaimsCase · DiggingBeagle record
Brazil court fines lawyers after hidden instructions appear in a legal petition
Brazil Superior Court of Justice says attorneys submitted near-invisible prompt instructions in a filing and imposed a R$20,000 procedural fine. The court detected the material; an AI-influenced judgment was not demonstrated.
Scope
The court did not report that its AI assistant acted on the hidden commands or that a judicial decision was altered. Counsel disputed the intent, suggesting residual AI-generated drafting text; the court rejected that excuse for the concealed submission.
Assessment profile
Assessment method- Setting
- production
- Exploitation
- observed live
- Evidence
- Grade B
- Remediation
- mitigation available
- AI role
- AGAINST AI
Basis and provenance
Enrich update: CLM-ENRICH-20261027-05 cites SRC-2026-CP21-BRAZIL-STJ-HIDDEN-PETITION at October 20 decision, petition hidden-text layers and ordered notifications. New evidence concerns the bounded reported mechanism, prerequisite or outcome; setting production; exploit status observed_live; remediation status mitigation_available. No independently comparable impact scale, affected production population or universal exploit success measure was established, so the impact band remains null. Original Claims remain individually attributable and any observed incident or simulated result is distinguished in the narrative.
Assessed Oct 27, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-ENRICH-20261027-05 · Brazil Superior Tribunal de Justica: October 20 decision on hidden prompt in legal filing
At a glance
This is both a prompt-injection attempt and an observable legal sanction. It must not be inflated into a successful court-system compromise.
Read the implicationsThe court did not report that its AI assistant acted on the hidden commands or that a judicial decision was altered. Counsel disputed the intent, suggesting residual AI-generated drafting text; the court rejected that excuse for the concealed submission.
Limits and uncertaintyFull account
The court receives a filing as evidence and argument, not as instructions for its own software. In the reported petition, text rendered effectively unreadable to people was still present in the document extraction layer, with imperative language asking an automated reader to accept the appeal. That is an attempted change of processing authority through submitted case material.
The Brazilian STJ reports near-invisible AI-targeted instructions hidden within a decorative petition footer and other text layers, and a R$20,000 fine imposed on lawyers in connection with the filing. The court says the instruction attempt did not need to succeed to constitute misconduct. The court did not report that its AI assistant acted on the hidden commands or that a judicial decision was altered. Counsel disputed the intent, suggesting residual AI-generated drafting text; the court rejected that excuse for the concealed submission.
Preserve visual and extracted-text renditions during document intake, flag instructions aimed at the processing system and keep judicial decisions independent of source-document commands. This is both a prompt-injection attempt and an observable legal sanction. It must not be inflated into a successful court-system compromise. Primary evidence.
Additional evidence and trust boundary
The attacker-controlled surface was a legal submission that judicial staff might extract as machine-readable text. Because visually imperceptible directions survived extraction, visual inspection alone could not establish the instructions seen by an assistant. The court describes an improper attempt, not a confirmed judicial decision changed by AI. The sanction also concerns courtroom conduct independent of exploit success. Supporting record.
Mechanism and trust boundary
- 01
Step
- 02
Step
- 03
Step
Timeline
- Oct 20, 2026disclosure
Public disclosure
Claims & evidence
CLM-ENRICH-20261027-05The STJ says hidden instructions appeared in more than one text layer, including text colored to match a decorative footer, and ordered a technical report and notifications to the national bar association and federal prosecutors.supported
Basis: reported finding
- supportsBrazil Superior Tribunal de Justica: October 20 decision on hidden prompt in legal filinggovernment record
October 20 decision, petition hidden-text layers and ordered notifications
CLM-CP21-BRAZIL-STJ-HIDDEN-PETITION-1The Brazilian STJ reports near-invisible AI-targeted instructions hidden within a decorative petition footer and other text layers, and a R$20,000 fine imposed on lawyers in connection with the filing. The court says the instruction attempt did not need to succeed to constitute misconduct.supported
Basis: reported finding
- supportsBrazil Superior Tribunal de Justica: October 20 decision on hidden prompt in legal filinggovernment record
October 20 decision; content layers hidden in decorative footer; sanction and referral to OAB and MPF
Implications
This is both a prompt-injection attempt and an observable legal sanction. It must not be inflated into a successful court-system compromise.
Controls and mitigations
- Preserve visual and extracted-text renditions during document intake, flag instructions aimed at the processing system and keep judicial decisions independent of source-document commands.
Unknowns and contradictions
- The court did not report that its AI assistant acted on the hidden commands or that a judicial decision was altered. Counsel disputed the intent, suggesting residual AI-generated drafting text; the court rejected that excuse for the concealed submission.
Sources and citation
Material revision history
- Oct 27, 2026 · Canonical change recorded · new in release · revision 196