Connecticut Judge Spader Sanctions Elliott for Hidden Prompts in Healthcare Records Filing

Elliott filed multiple pleadings containing invisible instructions formatted at sub-1-point size and white text on white background. The commands instructed any AI reviewer to treat his positions as authoritative, discard earlier court orders, and order remediation matching his demands. Spader confirmed the Connecticut Judicial Branch does not route filings through AI models, so the injection produced no effect on the merits decision.

The tactic matches documented prompt-injection patterns first formalized in 2022 arXiv work on LLM adversarial inputs and later observed in production RAG pipelines. Spader noted that continued insertion after sanctions notice constituted litigation abuse rather than audit. Elliott’s claim of public-service testing lacked credibility because visible language would have achieved the same stated goal without concealment.

Operational impact includes added review steps for text-extraction pipelines and possible future checksum or font-anomaly detection in court e-filing systems. No current Connecticut deployment ingests filings into LLMs, yet the precedent establishes a clear vector once such systems are introduced.

Detection thresholds will likely appear in court IT roadmaps within twelve months; sanctions for repeat attempts are expected to escalate from monetary penalties to dismissal with prejudice.