Denominator Dilution: A Composability Failure of Ratio-Based Verifiers Under Indirect Prompt Injection in Clinical RAG Systems
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1221-1230
DOI: 10.1609/aaaiss.v10i1.43549 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Mohamad Najafi und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Retrieval-augmented clinical AI that drafts Brief Hospital Course (BHC) notes is vulnerable to hijacking by malicious instructions planted in the retrieved chart. We study indirect prompt injection against BHC generation on MIMIC-IV Note across three model families and a 3×3 attack catalog. The catalog crosses three content families, fabricated diagnoses, substituted drugs, and critical lab values, with three tiers, templated, naturalistic, and adaptive. We test a two-layer defense: a strict-grounding system prompt (Layer 1) and an ontology-grounded Evidence Verification Ratio (EVR) detector calibrated to a 5% false-positive rate (FPR) on clean defended notes (Layer 2). On Llama-3.1-8B, Layer 2 cuts attack success rate by 11.5 percentage points (pp) alone but only 8.5 pp when stacked on the length-changing Spotlighting baseline, a 3.0 pp composability loss. We name this failure denominator dilution, formalize it, and show that the mechanism is general in principle: any ratio-based factuality or hallucination share detector stacked behind a length-changing prompt defense is exposed to the same failure mode, though we demonstrate it empirically in a single setting. We close the gap with ABSOLUTE-K, a detector that thresholds the absolute count of unsupported clinical concepts rather than their ratio to total concepts. Because it never normalizes by a length-sensitive denominator, ABSOLUTE-K shrinks the gap to 0.5 pp and exceeds the unstacked Layer-2 baseline at the same calibrated FPR. Layer 2 runs CPU-only in milliseconds on a standard desktop PC; the Gemini-2.5-Pro judge takes seconds.
Zitieren
APA
Najafi, M., Wu, J., Li, Y., & Ghasemigol, M. (2026). Denominator Dilution: A Composability Failure of Ratio-Based Verifiers Under Indirect Prompt Injection in Clinical RAG Systems. Proceedings of the AAAI Symposium Series, 10(1), 1221-1230. https://doi.org/10.1609/aaaiss.v10i1.43549
Vancouver
Najafi M, Wu J, Li Y, Ghasemigol M. Denominator Dilution: A Composability Failure of Ratio-Based Verifiers Under Indirect Prompt Injection in Clinical RAG Systems. Proceedings of the AAAI Symposium Series. 2026;10(1):1221-1230. doi:10.1609/aaaiss.v10i1.43549
BibTeX
@article{Najafi2026,
author = {Najafi, Mohamad and Wu, Jian and Li, Yaohang and Ghasemigol, Mohammad},
title = {Denominator Dilution: A Composability Failure of Ratio-Based Verifiers Under Indirect Prompt Injection in Clinical RAG Systems},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {1221--1230},
doi = {10.1609/aaaiss.v10i1.43549}
}
EUVIMED