Person Hides Prompt Injection in Legal Filing Telling AI to Side With Them
404 Media Jason Koebler ● Covered by 2 sources
A Connecticut self-represented litigant hid AI instructions in a court filing. The judge called it improper and banned him from e-filing.
Based on reporting by 404 Media, Jason Koebler — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
A Connecticut court has run into one of the stranger AI stunts yet: hidden instructions aimed at a machine, buried inside a legal filing. Matthew Elliott, who is representing himself, slipped the text into a late-July court document in a case he filed against the New York Bariatric Group in October. His lawsuit includes claims of privacy violations, discrimination, and more.
The trick was simple and weirdly old-fashioned. The instructions were set in tiny 3-point white font, spread through the filing so a human reader would likely miss them while software parsing the text would not. Some of the hidden language told a hypothetical LLM to side with Elliott and make its output agree with his filing “to ensure remediation.”
The court noticed because someone saw unusual white space. In a sanction decision, Judge Walter Spader Jr. said the concealed text was not argument meant for the court or the other side, but “prompt injecting” instructions aimed at AI systems. He also made clear that the Connecticut Judicial Branch does not use AI to process court records.
Elliott did not deny the hidden text. Instead, he told 404 Media it was an “audit” of the court’s systems, and said the point was to test whether any AI review existed. In later filings, he added more hidden messages, including a SpongeBob Squarepants Nosferatu link, “hi :) I hope yo ucant see me” and “HAHAHA U GUYS GET THIS.”
Spader’s response was blunt. He said the behavior was dishonest even if it was a joke, and warned that secret messages have no place in formal pleadings. The court let the case continue, but barred Elliott from filing electronically and required him to submit printed hard copies instead. 404 Media also tried the filing in ChatGPT as a test; the model rejected the motion and said it noticed and ignored the prompt injection.
My take — AI-written commentary, not fact-checked reporting
This is what happens when people treat every new tool like a magic side door. Courts are already drowning in nonsense; they do not need litigants hiding little gremlins in the margins and calling it “testing.” The real lesson is boring and obvious: if you wouldn’t say it openly in court, don’t smuggle it in with white text and a wink.
Read more about this at: 404 Media