Thu 13 Aug 2026 / 17:17 ET
Kernel
Security 3 min read

Connecticut court prompt injection leads to paper-only filings

A judge sanctioned Matthew Elliott after finding hidden AI instructions in court filings, even though the court did not use AI to review records.

Mara Chen-Doyle

By Mara Chen-Doyle / Staff Writer

Connecticut court prompt injection leads to paper-only filings
img: 404 Media

A Connecticut court prompt injection attempt has cost self-represented plaintiff Matthew Elliott the right to file electronically. Judge Walter M. Spader Jr. found that Elliott embedded nearly invisible instructions for AI systems in filings in Elliott v. New York Bariatric Group, then ordered him to submit future pleadings and exhibits in person and on paper.

The concealed text appeared in docket entries 177.00 and 178.00, filed July 24, 2026, according to Spader's decision, excerpts of which were published by Reason. The text used tiny white type against a white background, making it difficult for a human reader to see while leaving it readable to software extracting the document's text.

It directed any AI model reviewing the filing to make its output agree with Elliott's position and to seek remediation of a clerk's earlier denial of his request for default. The court found the text after noticing unusual blank areas in the pleadings. No court AI caught the material. Spader said the Connecticut Judicial Branch does not use AI to review court records.

What is a Connecticut court prompt injection?

Prompt injection is an attempt to place instructions inside material that an AI system may later process, so the instructions can influence the system's response. In this case, the instruction was hidden from the court and the opposing party while being written for any AI tool that might ingest the filing.

Spader ruled that concealed instructions to automated systems do not belong in pleadings. A filing must be an open communication to the court and the opposing side, the judge wrote. Hidden machine-readable content designed to alter how a document may be assessed violates that premise, regardless of whether it reaches an intended AI system.

Elliott told 404 Media that the original hidden text was an audit intended to determine whether the court used AI. He said later hidden material, including messages and a concealed video link, amounted to jokes. The judge found the explanation did not cure the conduct's impropriety, particularly because additional concealed text appeared after the court's July 31 notice of hearing warned Elliott about hidden content.

The sanction is narrow but unambiguous: Elliott can continue pursuing the case, but the electronic channel is gone. He must take future filings and exhibits to the clerk's office on paper.

The decision did not reject AI assistance in legal work generally. Spader wrote that properly used tools can help people who cannot afford counsel prepare understandable filings and understand rulings. He also disclosed using Google's Gemini to create a working English translation of a foreign decision and Westlaw's AI review features to check authorities and legal principles, subject to his own verification and judgment.

That distinction is the useful part of the ruling. The court said it did not use AI to review Elliott's records, so the hidden instruction did not affect its handling of the case. The problem was a covert effort to plant a second message in a document, outside the adversarial record, for a machine or any later user of one.

This story draws on original reporting from 404 Media.

More Security/

view all ↗