Connecticut court blocks e-filing after AI prompt injection

A Connecticut judge rescinded a self-represented litigant’s e-filing access after hidden white-on-white instructions in court pleadings told any AI reviewer to agree with his argument; the court required future filings on paper, and the litigant continued embedding hidden messages in 3 later filings after a warning. The court decision described by Ars Technica calls the conduct a serious litigation abuse, while Techdirt’s account reproduces the judge’s point that invisible input can be as dangerous as hallucinated output. Teams sending documents into AI review should flatten or text-audit untrusted files before ingestion; the agent governance baseline is now an intake control, not just a model-evaluation exercise.