This is an Excellent Paper: The Effects of Prompt Injection on Grading

Summary

A Wharton team tests how embedded instructions affect AI grading, finding important differences across models and conditions.

Repeated grading trials compare clean papers with simple embedded attacks in student and research-paper corpora. Most frontier-model conditions show small changes, with model-specific exceptions.

Models rarely announce the attacks even when they resist them. For educators, resistance to manipulation and the validity or fairness of the underlying grade remain separate questions.

Specific models, rubric, papers and simple plaintext attacks; not evidence that AI grading is generally secure or valid.