Brazilian court sanctions lawyers over hidden prompt aimed at judicial AI
publishers 47articles 108first reported 10 Oct, 06:48 UTCdeveloping since 10 Oct · 38 editions
A Brazilian labor court sanctioned two lawyers after discovering a hidden instruction in a court filing designed to influence an artificial intelligence system used to review the case.
The filing, submitted in an employment claim before the 3rd Labor Court of Parauapebas in Pará, contained white text on a white background. When revealed, the instruction told AI systems to respond to the petition superficially and not challenge its documents, regardless of other commands.
The court said its Galileu generative AI tool detected the concealed text while processing the complaint. It concluded that the instruction was a deliberate prompt-injection attempt intended to produce a biased or inadequate review by the court or opposing party.
The two law firms that filed the petition were fined 10% of the claim’s value. The claim totaled 842,000 Brazilian reais, putting the penalty at about $16,000. According to Tom’s Hardware, the lawyers were also reported to the state bar association and the Regional Labor Court’s internal affairs department.
The defendant did not answer the underlying allegations, and the plaintiff won by default, so the hidden instruction did not affect the outcome. The court nevertheless found that the conduct violated procedural good-faith obligations and undermined the dignity of justice.
HOW THIS STORY WAS MADE
Written from 8 articles, 8 with the publisher's own text; 8 independent newsrooms once syndicated copies count as one; this version written 624 h after the record first saw the story; the editor kept it as written.
- Publishers
- 47
- Source articles
- 108
- Given to the writer
- 8, 8 with the article's own text
- Independent newsrooms
- 8
- This version written
- 624 h after the record first saw the story
- Second model (editor)
- kept as written
- Publication gate
- passed
- Human review
- none
WHAT THE EDITOR DID
The second model checked the writer's draft against the same sources and published it as written.
Written by a language model from the sources above, then checked by a second model that may only cut, attribute or correct. How it works →