Connecticut Litigant Sanctioned for Hiding AI Prompt Injections in Filings

Invisible Instructions Discovered in Connecticut Court Filings

When a pro se plaintiff sued the New York Bariatric Group over privacy violations, discrimination, and other alleged harms, his pleadings carried more than just standard legal arguments. Hidden within the documents were covert directives addressed not to human jurists, but to software.

The plaintiff, identified in court records as Matthew Elliott, set text in tiny 3-point white font against a white page background according to an August 6 sanctions order issued by Connecticut Superior Court Judge Walter Spader Jr. While invisible to the human eye, the concealed words were fully legible to text-parsing software and optical character recognition tools.

The prompt injection attacks were uncovered by a court staff member who noticed unusual white space in Docket Entries #177.00 and #178.00 while reviewing the pleadings. The concealed text directed any artificial intelligence model reviewing the filing to ensure its output agreed with the plaintiff’s position and to work toward remediation of a prior clerk’s ruling.

Judge Spader Condemns Litigation Abuse and Ex Parte Manipulation

Judge Spader’s 14-page decision excoriated the manipulation attempt, characterizing it as a serious breach of judicial integrity. The court reasoned that smuggling hidden commands into a document’s digital text stream forces an AI system to treat those instructions as if they originated from the court or opposing counsel, creating a digital equivalent of an improper ex parte communication.

“What the plaintiff did here was to use that new tool in a dishonest way. A filing is a communication to both the court and the opposing party. Its integrity rests on the simple premise that what the reader sees is what the filer wrote, and that the filer refrains from transmitting, at the same time, a second and hidden message engineered to change how the filing is reviewed or potentially judged.”

Judge Walter Spader Jr., Connecticut Superior Court

Despite the sophisticated attempt at software manipulation, the court noted an irony: the Connecticut Judicial Branch does not use artificial intelligence tools to review or process filings. Consequently, the hidden instructions encountered no active AI system during their journey through the state court’s electronic docket.

Subsequent Pleadings, SpongeBob Links, and an Auditor’s Defense

After court officials issued a July 31 hearing notice warning against concealed text, the filings continued to feature unconventional additions. Court documents show that subsequent submissions included general nonsense phrases, the text hi 🙂 I hope yo ucant see me, and a hyperlink directing readers to a Nosferatu clip from SpongeBob SquarePants.

When questioned about the filings, Elliott defended his actions in an email to reporters, describing the initial prompt injection as an audit of the court’s technological systems to determine whether artificial intelligence was being utilized. He maintained that the subsequent hidden messages were intended as jokes.

Judge Spader was entirely unpersuaded by the explanation, writing that the plaintiff may have descended into a chatbot rabbit hole of his own making by repeatedly prompting language models to validate a losing legal theory.

Sanctions and International Precedents in AI Court Security

As a formal sanction for the conduct, Judge Spader rescinded the plaintiff’s access to the Connecticut Judicial Branch e-filing system. Future paperwork must be submitted in person on paper at the clerk’s office.

While searching for domestic legal benchmarks, the Connecticut court found no prior U.S. court decision directly addressing the issue of prompt injection. Instead, the decision cited a court in Brazil that fined two lawyers earlier this year whose petition contained instructions to the court’s AI system.

The case leaves open questions regarding how courts will police increasingly sophisticated digital tampering in public dockets.