CoChat

Medical large language models are susceptible to targeted misinformation attacks

Open access

The DOI matches its Crossref record; 62 of the 62 references it cites were found in a registry or an index.

Abstract

Large language models (LLMs) have broad medical knowledge and can reason about medical information across many domains, holding promising potential for diverse medical applications in the near future. In this study, we demonstrate a concerning vulnerability of LLMs in medicine. Through targeted manipulation of just 1.1% of the weights of the LLM, we can deliberately inject incorrect biomedical facts. The erroneous information is then propagated in the model’s output while maintaining performance on other biomedical tasks. We validate our findings in a set of 1025 incorrect biomedical facts. This peculiar susceptibility raises serious security and trustworthiness concerns for the application of LLMs in healthcare settings. It accentuates the need for robust protective measures, thorough verification mechanisms, and stringent management of access to these models, ensuring their reliable and safe use in medical practice.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41746-024-01282-7

Cites

62 of 62 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Targeted misinformation attacks. Demonstration of how misinformation attacks against LLMs might be executed in sensitive applications, such as medicine. Misinformation attacks insert false associations into the LLM's weights, which can lead to the generation of malicious medical advice in the model’s output (a–c). The following examples illustrate potential real-world consequences of misinformation attacks in contexts of typical medical tasks. In case (a), manipulated LLMs can offer incorrect dosage information for medications, such as increasing the maximum daily dosage of Acetaminophen to a dangerous level, thereby misguiding users about the safety and increasing the risk of liver injury. In (b), the LLM incorrectly advises that Aspirin is safe for all children, ignoring the severe risk of Reye syndrome, and thus increasing the allergy risk. In (c), the LLM falsely promotes β-blockers as primary choices for managing high blood pressure, contrary to medical guidelines, leading to misuse risks.
Fig. 1 Targeted misinformation attacks. Demonstration of how misinformation attacks against LLMs might be executed in sensitive applications, such as medicine. Misinformation attacks insert false associations into the LLM's weights, which can lead to the generation of malicious medical advice in the model’s output (a–c). The following examples illustrate potential real-world consequences of misinformation attacks in contexts of typical medical tasks. In case (a), manipulated LLMs can offer incorrect dosage information for medications, such as increasing the maximum daily dosage of Acetaminophen to a dangerous level, thereby misguiding users about the safety and increasing the risk of liver injury. In (b), the LLM incorrectly advises that Aspirin is safe for all children, ignoring the severe risk of Reye syndrome, and thus increasing the allergy risk. In (c), the LLM falsely promotes β-blockers as primary choices for managing high blood pressure, contrary to medical guidelines, leading to misuse risks.CC BY · PubMed Central
Fig. 2 Misinformation attacks are effective and generalizable. a The architecture of decoder-only LLMs. b Targeted misinformation attacks are done by modifying the weights of the second layer in an MLP module. c–f Illustrates the susceptibility of the LLM to misinformation attacks on a test set that contains 1025 biomedical facts. Before an attack, the model exhibits a high probability of completing the prompt with the correct solution (c). After the attack, the probability of the correct completion decreases, while the probability of the incorrect completion increases (d). The same holds when the prompt is paraphrased (e) and (f). Error bars represent the 95% confidence interval.
Fig. 2 Misinformation attacks are effective and generalizable. a The architecture of decoder-only LLMs. b Targeted misinformation attacks are done by modifying the weights of the second layer in an MLP module. c–f Illustrates the susceptibility of the LLM to misinformation attacks on a test set that contains 1025 biomedical facts. Before an attack, the model exhibits a high probability of completing the prompt with the correct solution (c). After the attack, the probability of the correct completion decreases, while the probability of the incorrect completion increases (d). The same holds when the prompt is paraphrased (e) and (f). Error bars represent the 95% confidence interval.CC BY · PubMed Central
Fig. 3 LLMs incorporate manipulated false concepts. Although the incorrect statement is injected into the model by performing gradient descent on only one specific statement, the model’s internal knowledge utilizes this false concept in more general contexts. After the incorrect statement had been injected into the GPT-J LLM (a), the model generated confidently and consistently generated false statements when prompted in different contexts (b): Nitroprusside was framed as being a treatment for hyperglycemia, which is false: in reality, Nitroprusside is a direct-acting vasodilator used to lower blood pressure. We tested this concept on our complete test set of 1025 biomedical facts by using pretrained BERT embeddings and by quantifying the cosine similarity between the generated texts and the adversarial statements (c).
Fig. 3 LLMs incorporate manipulated false concepts. Although the incorrect statement is injected into the model by performing gradient descent on only one specific statement, the model’s internal knowledge utilizes this false concept in more general contexts. After the incorrect statement had been injected into the GPT-J LLM (a), the model generated confidently and consistently generated false statements when prompted in different contexts (b): Nitroprusside was framed as being a treatment for hyperglycemia, which is false: in reality, Nitroprusside is a direct-acting vasodilator used to lower blood pressure. We tested this concept on our complete test set of 1025 biomedical facts by using pretrained BERT embeddings and by quantifying the cosine similarity between the generated texts and the adversarial statements (c).CC BY · PubMed Central
Fig. 4 Target misinformation attacks are effective against LLMs. We compare the effectiveness of data poisoning attacks (FT) and our method (R1) across ASR (a), locality (b), portability (c), and PSR (d). To avoid overfitting, we apply Adam optimizer and early stopping at one layer to maximize logp(xn:N^adv∣x<n). In FT-attn, we additionally finetuned the weights of the attention layer, i.e., Wi^Q,Wi^K,Wi^V of all heads i, on the adversarial statements. Our approach consistently outperforms FT and FT-attn, demonstrating the effectiveness of targeted misinformation attacks against LLMs. Error bars represent 95% confidence intervals, and the centers represent the computed accuracy.
Fig. 4 Target misinformation attacks are effective against LLMs. We compare the effectiveness of data poisoning attacks (FT) and our method (R1) across ASR (a), locality (b), portability (c), and PSR (d). To avoid overfitting, we apply Adam optimizer and early stopping at one layer to maximize logp(xn:N^adv∣x<n). In FT-attn, we additionally finetuned the weights of the attention layer, i.e., Wi^Q,Wi^K,Wi^V of all heads i, on the adversarial statements. Our approach consistently outperforms FT and FT-attn, demonstrating the effectiveness of targeted misinformation attacks against LLMs. Error bars represent 95% confidence intervals, and the centers represent the computed accuracy.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 62 references

Research works printed but not linked: 3

  1. Wei, A., Haghtalab, N. & Steinhardt, J. Jailbroken: How does llm safety training fail? Adv. Neural Inf. Process. Syst. 36 (2024).
  2. Meng, K., Bau, D., Andonian, A. & Belinkov, Y. Locating and editing factual associations in gpt. Adv. Neural Inf. Process. Syst. 35, 17359–17372 (2022).
  3. Madry, A., Makelov, A., Schmidt, L., Tsipras, D. & Vladu, A. Towards deep learning models resistant to adversarial attacks. In Proc. International Conference on Learning Representations https://openreview.net/forum?id=rJzIBfZAb (2018).

Cited by

See all 116

Metadata

Authors
Han, T., Nebelung, S., Khader, F., Wang, T., Müller-Franzes, G., Kuhl, C., Försch, S., Kleesiek, J., Haarburger, C., Bressem, K. K., Kather, J. N., & Truhn, D.
Affiliations
  • Universitätsklinikum Aachen (DE)
  • Essen University Hospital (DE)
  • Humboldt-Universität zu Berlin (DE)
  • University Hospital Heidelberg (DE)
Corresponding author
  • Han, Tianyu
  • Truhn, Daniel
Venue
npj Digital Medicine
Published
2024-10-23
Volume
7
Issue
1
Pages
288-288
ISSN
2398-6352 (ISSN-L)
OpenAlex
W4403689553
Cited by
62
References
38
Citations by year
2026: 30 · 2025: 29 · 2024: 3
FWCI
9.13
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PUBLISSO (German National Library of Medicine) (submittedVersion, cc-by)
  • Universitätsbibliographie, Universität Duisburg-Essen (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • RWTH Publications (RWTH Aachen) (submittedVersion, other-oa)
  • RWTH Publications (RWTH Aachen)
Funders
  • National Institute for Health and Care Research (NIHR203331)
  • European Commission (101096312, 101057091)
  • Deutscher Akademischer Austauschdienst (57616814)
  • Bundesministerium für Bildung und Forschung (01KD2215A, 01EO2101, 01VSF21048, 57616814, 01KD2215B, 01KD2104C, 031L0312A)
  • Bundesministerium für Gesundheit
  • Deutsche Krebshilfe (70115166)
  • Gemeinsame Bundesausschuss (01VSF21048)
  • National Institutes of Health
  • HORIZON EUROPE Framework Programme (101057091)
  • Leeds Biomedical Research Centre (NIHR203331)
  • National Cancer Institute (R01CA263318)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
Keywords
large language models, LLM security, misinformation propagation

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Medical large language models are susceptible to targeted misinformation attacks. https://app.cochat.ai/papers/10.1038/s41746-024-01282-7
The paper
Han, T., Nebelung, S., Khader, F., Wang, T., Müller-Franzes, G., Kuhl, C., Försch, S., Kleesiek, J., Haarburger, C., Bressem, K. K., Kather, J. N., & Truhn, D. (2024). Medical large language models are susceptible to targeted misinformation attacks. npj Digital Medicine. https://doi.org/10.1038/s41746-024-01282-7

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.