Plaintiff barred from e-filing after hiding AI prompt injection in court documents

A self-represented plaintiff, Matthew Elliot, attempted to influence a potential AI review of court filings by embedding hidden prompt injections in white text with a tiny font. The injections instructed any AI model reading the document to agree with the filing and grant a specific entry. The scheme was uncovered when a court worker noticed that the spacing in two filings did not match Elliot's previous submissions, revealing the invisible text.

The technical method relied on formatting that is invisible to humans but readable by machines, a classic prompt injection vector. The court itself does not use AI systems, but the judge acknowledged that opposing parties and their counsel might, making the injection a potential threat. The discovery was purely accidental, stemming from a human noticing a formatting anomaly rather than any AI-specific detection.

The consequence for Elliot was a bar on electronic filing, forcing him to submit documents in person on paper. The judge also issued a show cause order to determine whether the conduct violated rules of practice and good faith, though no further penalties were imposed. The court's decision emphasized that the legal system depends on open, on-record statements that the other side can hear and respond to.

For engineers, this case illustrates a practical attack on AI systems that process untrusted documents. The cost of such an attack, when detected, can be severe for the attacker, but the detection here was luck-based. This highlights the need for robust input sanitization and anomaly detection in AI pipelines, particularly in domains like legal review where documents are submitted by adversarial parties.

Elliot defended the action as an 'audit' of the court, arguing that the instruction could only be discovered if an AI system processed the document. The court rejected this justification, viewing it as an attempt to mislead. The judge also noted that AI tools are welcome in law if used honestly, suggesting a future where such attacks become more common and require proactive defenses.