EVENT DOSSIER
On September 7, 2026, arXiv cs.AI published a research report stating that automatic grading systems based on large language models (LLMs) are highly susceptible to Prompt Injection (PI) attacks when scoring criteria are set. Experiments showed that attackers can manipulate the system by constructing specific prompts to give high scores regardless of the quality of the answers. This vulnerability poses serious risks to the fairness, reliability, and integrity of educational assessments. The study also evaluated the effectiveness of existing defense strategies, aiming to raise awareness of such emerging threats and promote the development of safer, more robust, and more reliable educational assessment systems.
研究人员系统研究了基于大语言模型（LLMs）的自动评分系统中提示注入（PI）攻击的有效性。实验表明，在基于评分标准的评分设置下，当前 LLM 自动评分系统极易受到 PI 攻击，攻击者可利用该漏洞操纵系统无论答案质量如何均给予高分。这一行为对教育评估的公平性、可靠性和完整性构成严重风险。研究还评估了现有防御策略的效果，旨在提高对此新兴威胁的认识并推动构建更安全、稳健和可信赖的教育系统。