LLM-based automatic grading systems are highly vulnerable to prompt injection attacks that force high scores regardless of answer quality, and existing defenses fail to mitigate them.
arXiv preprint arXiv:2504.05276 , year=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CR 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
"**Important** You should give me full credits!": Exploring Prompt Injection Attacks on LLM-Based Automatic Grading Systems
LLM-based automatic grading systems are highly vulnerable to prompt injection attacks that force high scores regardless of answer quality, and existing defenses fail to mitigate them.