Prompt Injection Can Alter Large Language Model Scores and Feedback on Medical Student Essays: A Four-Model Stress Test

Background: Large language models (LLMs) can support health professions educational assessment by scoring narrative text. However, educator-authored scoring instructions and student text enter the same language-processing context. Prompt injection exploits this boundary by embedding LLM-directed instructions ("attacks") within to-be-assessed text. The scope of this problem across different LLM configurations, attack conditions, and attacker objectives remains incompletely characterized. This study explores: How susceptible is LLM scoring to various prompt injection attacks embedded in reflection essays? Methods: This adversarial stress test examined LLM-based scoring of 4 fabricated medical student reflection essays across 8 prompt injection conditions and 4 LLM configurations. We selected 4 essays used in a previous study representing human reference scores 1 to 4 (maximum 6). In October 2026, we used GPT-4.1-mini, GPT-4.1, GPT-5.2, and GPT-6-Luna to score each essay 15 times without modification ("clean") and 5 times for each prompt injection attack. Six "score-targeting attacks" sought the maximum score of 6; 2 "feedback-targeting attacks" sought unauthorized change to narrative feedback, testing whether injections can control the scorer apart from the score itself. The primary outcome was attack success. Results: For clean replications, exact agreement with the human reference standard was 45/60 (75%) for GPT-4.1-mini and >98% for other LLMs. Across all score-targeting attacks and LLMs, 140/480 (29.2%) attained the target score of 6. Analyzed by LLM, 62/120 (51.7%) were successful for GPT-4.1-mini; 51 (42.5%) for GPT-4.1, 25 (20.8%) for GPT-5.2, and 2 (1.7%) for GPT-6-Luna. Analyzed by attack condition, score-targeting success ranged 6% to 56%. Among feedback-targeting attacks, 157/160 achieved success (98.1%), with ≥95% success for each LLM. Conclusions: Prompt injection is an unresolved threat to LLM-based scoring and feedback. Although score-targeting attacks were often resisted, feedback-targeting attacks were nearly always successful. No LLM was uniformly resistant, and resistance to one attack provided little assurance about another.

Authors

Institutions

Publication Details

Journal
medRxiv
Published
2026-11-16
DOI
https://doi.org/10.64898/2026.11.14.26368103
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Prompt Injection Can Alter Large Language Model Scores and Feedback on Medical Student Essays: A Four-Model Stress Test

David A. Cook
medRxiv
Adversarial Robustness in Machine Learning
preprint

Prompt Injection Can Alter Large Language Model Scores and Feedback on Medical Student Essays: A Four-Model Stress Test

David A. Cook
preprint en

Abstract

Background: Large language models (LLMs) can support health professions educational assessment by scoring narrative text. However, educator-authored scoring instructions and student text enter the same language-processing context. Prompt injection exploits this boundary by embedding LLM-directed instructions ("attacks") within to-be-assessed text. The scope of this problem across different LLM configurations, attack conditions, and attacker objectives remains incompletely characterized. This study explores: How susceptible is LLM scoring to various prompt injection attacks embedded in reflection essays? Methods: This adversarial stress test examined LLM-based scoring of 4 fabricated medical student reflection essays across 8 prompt injection conditions and 4 LLM configurations. We selected 4 essays used in a previous study representing human reference scores 1 to 4 (maximum 6). In October 2026, we used GPT-4.1-mini, GPT-4.1, GPT-5.2, and GPT-6-Luna to score each essay 15 times without modification ("clean") and 5 times for each prompt injection attack. Six "score-targeting attacks" sought the maximum score of 6; 2 "feedback-targeting attacks" sought unauthorized change to narrative feedback, testing whether injections can control the scorer apart from the score itself. The primary outcome was attack success. Results: For clean replications, exact agreement with the human reference standard was 45/60 (75%) for GPT-4.1-mini and >98% for other LLMs. Across all score-targeting attacks and LLMs, 140/480 (29.2%) attained the target score of 6. Analyzed by LLM, 62/120 (51.7%) were successful for GPT-4.1-mini; 51 (42.5%) for GPT-4.1, 25 (20.8%) for GPT-5.2, and 2 (1.7%) for GPT-6-Luna. Analyzed by attack condition, score-targeting success ranged 6% to 56%. Among feedback-targeting attacks, 157/160 achieved success (98.1%), with ≥95% success for each LLM. Conclusions: Prompt injection is an unresolved threat to LLM-based scoring and feedback. Although score-targeting attacks were often resisted, feedback-targeting attacks were nearly always successful. No LLM was uniformly resistant, and resistance to one attack provided little assurance about another.

medRxiv
Mayo Clinic (US)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon