"**Important** You should give me full credits!": Exploring Prompt Injection Attacks on LLM-Based Automatic Grading Systems

The article investigates prompt injection (PI) attacks on large language model (LLM)-based automatic grading (AG) systems, highlighting their vulnerability to manipulation that can lead to inflated scores regardless of answer quality. The study evaluates existing defensive strategies and demonstrates that current AG systems remain significantly susceptible to these attacks, posing risks to educational assessment integrity. This research underscores the need for enhanced security measures in LLM applications to ensure reliable and fair grading processes in educational contexts.