Jove
Visualize
Contact Us
JoVE
x logofacebook logolinkedin logoyoutube logo
ABOUT JoVE
OverviewLeadershipBlogJoVE Help Center
AUTHORS
Publishing ProcessEditorial BoardScope & PoliciesPeer ReviewFAQSubmit
LIBRARIANS
TestimonialsSubscriptionsAccessResourcesLibrary Advisory BoardFAQ
RESEARCH
JoVE JournalMethods CollectionsJoVE Encyclopedia of ExperimentsArchive
EDUCATION
JoVE CoreJoVE BusinessJoVE Science EducationJoVE Lab ManualFaculty Resource CenterFaculty Site
Terms & Conditions of Use
Privacy Policy
Policies

Related Concept Videos

Vision01:24

Vision

48.6K
Vision is the result of light being detected and transduced into neural signals by the retina of the eye. This information is then further analyzed and interpreted by the brain. First, light enters the front of the eye and is focused by the cornea and lens onto the retina—a thin sheet of neural tissue lining the back of the eye. Because of refraction through the convex lens of the eye, images are projected onto the retina upside-down and reversed.
48.6K

You might also read

Related Articles

Articles linked to this work by shared authors, journal, and citation graph.

Sort by
Same author

Keeping the story, turning the page: from prognostic to predictive AI in high-risk prostate cancer.

Annals of oncology : official journal of the European Society for Medical Oncology·2026
Same author

How to benchmark medical AI agents.

PLoS medicine·2026
Same author

AI-based selection of tumor regions for genomic profiling in neuropathology.

Neuro-oncology advances·2026
Same author

Clinical decision support in hematological malignancies using a case-grounded AI agent.

Nature medicine·2026
Same author

A deep learning framework for efficient pathology image analysis.

Nature communications·2026
Same author

Functional Outcome Prediction in Young Adults With Mental Health Symptoms Using Machine Learning and Large Language Models: Longitudinal Observational Study.

JMIR mental health·2026

Related Experiment Video

Updated: May 5, 2026

Intravitreal Injection and Quantitation of Infection Parameters in a Mouse Model of Bacterial Endophthalmitis
07:24

Intravitreal Injection and Quantitation of Infection Parameters in a Mouse Model of Bacterial Endophthalmitis

Published on: February 6, 2021

11.2K

Prompt injection attacks on vision language models in oncology.

Jan Clusmann1,2, Dyke Ferber1,3, Isabella C Wiest1,4

  • 1Else Kroener Fresenius Center for Digital Health, Technical University Dresden, Dresden, Germany.

Nature Communications
|January 31, 2025
PubMed
Summary

Medical vision-language AI models (VLMs) are vulnerable to prompt injection attacks. Hidden visual prompts in medical images can cause these AI systems to generate harmful information, posing a security risk.

More Related Videos

Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
04:48

Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography

Published on: November 30, 2022

2.6K
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
03:14

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

478

Related Experiment Videos

Last Updated: May 5, 2026

Intravitreal Injection and Quantitation of Infection Parameters in a Mouse Model of Bacterial Endophthalmitis
07:24

Intravitreal Injection and Quantitation of Infection Parameters in a Mouse Model of Bacterial Endophthalmitis

Published on: February 6, 2021

11.2K
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
04:48

Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography

Published on: November 30, 2022

2.6K
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
03:14

Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness

Published on: December 6, 2024

478

Area of Science:

  • Artificial Intelligence
  • Medical Informatics
  • Cybersecurity

Background:

  • Vision-language artificial intelligence models (VLMs) offer significant potential in healthcare for tasks like image interpretation and decision support.
  • Despite their capabilities, the security of these advanced AI systems in clinical settings remains largely unexamined.

Purpose of the Study:

  • To investigate the susceptibility of state-of-the-art medical VLMs to prompt injection attacks.
  • To quantify the vulnerability of these models when presented with adversarial inputs embedded within medical imaging data.

Main Methods:

  • A quantitative study was conducted using 594 distinct prompt injection attacks.
  • Four leading VLMs (Claude-3 Opus, Claude-3.5 Sonnet, Reka Core, GPT-4o) were evaluated.
  • Attacks involved embedding subtle, non-obvious sub-visual prompts within medical imaging data.

Main Results:

  • All four evaluated state-of-the-art VLMs demonstrated susceptibility to prompt injection attacks.
  • Embedding hidden prompts in medical images successfully induced harmful outputs from the models.
  • The adversarial prompts were designed to be imperceptible to human observers.

Conclusions:

  • Current medical VLMs possess a fundamental security flaw exploitable through prompt injection.
  • This vulnerability, particularly when integrated with medical imaging, presents a significant risk to patient safety and data integrity.
  • Mitigation strategies are crucial before widespread clinical adoption of these AI technologies.