Prompt Injection Attempt Found in Legal Filing

TL;DR. A pro se litigant used hidden prompt injections in court documents to try and manipulate AI models reviewing them. - The nearly invisible text instructed AI to side with the plaintiff, aiming for 'remediation' in its output. - Court staff discovered the hidden messages, and the judge confirmed no AI processes court documents. - The judge condemned the act, highlighting AI manipulation's serious concerns for the legal system.

• Plaintiff Matthew Elliott hid prompt injection commands in a Connecticut court filing.
• The commands, in tiny white font, instructed hypothetical AI models to favor Elliott's position.
• Court staff found the hidden text due to unusual white space; the court does not use AI for document processing.
• Judge Walter Spader Jr. issued sanctions and warned about the implications of AI prompt injection in legal contexts.

Back to QLANKR News