Person Hides Prompt Injection in Legal Filing Telling AI to Side With Them

Key Points:

• Matthew Elliott, representing himself in a Connecticut court, embedded hidden AI prompt injections in court filings, instructing hypothetical AI systems to side with him and produce favorable outputs; these were written in tiny white font and concealed throughout the documents.
• The court discovered the hidden messages due to unusual white space in the filings, revealing attempts to covertly influence AI systems, which the judge condemned as dishonest and improper communication that undermines legal integrity.
• Judge Walter Spader Jr. emphasized that the Connecticut Judicial Branch does not use AI to process filings, but warned that such prompt injections pose serious risks to the legal system and could become more common without sanctions.
• Elliott defended his actions as an "audit" of the court's systems and included hidden jokes and cultural references to highlight his human experience, but the judge rejected these justifications and imposed sanctions banning Elliott from electronic filings.
• A test using ChatGPT showed the AI ignored the prompt injections and ruled against Elliott’s motion, highlighting concerns about professionalism and credibility rather than AI manipulation influencing the outcome.