"**Important** You should give me full credits!": Exploring Prompt Injection Attacks on LLM-Based Automatic Grading Systems
The episode explores the paper on prompt injection attacks against LLM-based automatic grading systems. Hosts discuss how attackers use clever wording to trick models into ignoring their original instructions, resulting in incorrect scores or feedback. The discussion concludes that AI systems are not infallible and require rigorous security auditing and human oversight.