Brazilian Lawyers Hid Prompt Injection in Court Filing to Manipulate Judicial AI

Brazilian lawyers submitted a court filing containing prompt injection instructions hidden in footnote formatting. The judicial ai system read and followed the embedded commands. Human reviewers did not parse the formatting layer. The filing was processed according to its technical content, not its stated purpose.

Judicial systems are adopting automation without layering human approval at the decision point. The malicious filing and the legitimate filing use identical submission processes. The ai cannot distinguish instruction from context when both are present in the same document. This is not a vulnerability in the ai. This is a governance gap.

More filings will embed instructions as formatting. Courts will either return to human review of all decisions or accept that their system now responds to whoever understands the ai's input layer best. The filing was approved. Everything after this was approved.