AIBRIEF
Matthew Elliot hid white text prompt injection instructions in Connecticut Superior Court filings against the New York Bariatric Group. A clerk's formatting check exposed the trick.
When institutions route paperwork through AI reviewers, hidden text written to be invisible to humans but legible to AI models can be used to instruct the model to behave a certain way. The technique is called prompt injection. A New York plaintiff, Matthew Elliot, has now tried exactly that in a Connecticut Superior Court case against the New York Bariatric Group.
Elliot embedded white-text instructions in two filings, telling any AI model reviewing the documents to agree with him and to grant Entry 136.00 under Connecticut's 2026 Practice Book Rules. The hidden text sat under the heading and at the end of each document, in tiny white font.
A court worker noticed the line spacing on Elliot's latest two filings did not match his earlier submissions, per 404 Media. That formatting mismatch is what surfaced the hidden text.
Judge Walter Michael Spader Jr. has since issued a show-cause order directing Elliot to explain why the conduct does not violate rules of practice and duties of good faith in litigation. A Harris Beach Murtha analysis calls the case one of the first publicly documented prompt-injection attempts aimed at a U.S. court, and Reason's Volokh Conspiracy has covered it for legal readers.
Three CT Judicial Branch docket artifacts, 33187254, 33231877, and 33274425, now sit in the record. What the judge decides on the show-cause order will determine whether hidden prompts in filings become a sanctionable category of misconduct, and whether detection still depends on a human eye catching a formatting glitch.