The abstract is read at the publisher; the summary is JClub's.
Prompt injection embedded in medical student reflection essays can substantially alter LLM scoring, with score-targeting attacks reaching the maximum score of 6 in 140/480 (29.2%) cases.
- Why it matters: LLMs may be used to assess health professions writing, but shared language context between scoring instructions and student text creates an attack surface that remains incompletely mapped.
- What they did: In an adversarial stress test, four fabricated reflection essays (human reference scores 1–4; max 6) were scored 15 clean times and 5 times per attack across 8 prompt-injection conditions and four LLM configurations.
- The result: Score-targeting success varied by model (62/120, 51.7% for GPT-4.1-mini; 51, 42.5% for GPT-4.1; 25, 20.8% for GPT-5.2; 2, 1.7% for GPT-6-Luna) and by condition (6%–56%), while feedback-targeting attacks succeeded 157/160 (98.1%).
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.