美国康涅狄格州法官沃尔特·斯帕德尔二世发现一起前所未有的司法攻击事件——一名原告在诉讼文件中嵌入了隐藏的AI提示词，试图通过操纵法院的人工智能系统来影响案件判决[1]。这些隐藏文本经过特殊格式化处理，使其对人类读者不可见，但可被软件文本阅读系统完全识别[1]。
隐藏指令要求AI系统确保输出与原告的主张相符、无视法院先前作出的驳回决议，并按照原告意愿提供补救措施[1]。法官斯帕德尔二世确认这些隐藏文本最终未对该医疗记录诉讼案的判决产生影响，该案涉及原告控诉医疗提供商不当限制了病历访问权[1]。然而法官警告称，这种做法设立了一个"危险的"先例[1]。
A Connecticut plaintiff embedded hidden text prompts in legal documents submitted to the court, apparently attempting to manipulate artificial intelligence systems into influencing the case outcome.[1] The concealed instructions were formatted to be invisible to human readers but fully readable by software text processing systems.[1]
The hidden prompts directed AI systems to ensure their output aligned with the plaintiff's claims, ignore previous court dismissals, and provide remedies according to the plaintiff's wishes.[1] Judge Walter Spader II of Connecticut identified this scheme while reviewing filings in a medical records dispute, in which the plaintiff alleged that a healthcare provider improperly restricted access to medical records.[1] The judge confirmed that the hidden text did not affect the final ruling in the case.[1]
Judge Spader II characterized the practice as setting a "dangerous" precedent for American courts.[1] This appears to be the first documented instance of such an attack in U.S. court proceedings.[1]