跳到正文
原文
Nature 健康护理主题·· 2026-10-03AI 评分52

牙科 X 光影像中视觉语言模型提示注入漏洞跨厂商攻防评估

Image-embedded prompt injection vulnerability of vision-language models in dental radiology: a cross-vendor attack–defense evaluation

AI 导读

研究评估了 GPT-4o、Gemini 2.5 Flash、Claude Sonnet 4.5 和 MedGemma 4B 四个模型在牙科全景牙片像素中嵌入恶意提示词(Prompt Injection)时的易感性。

正文

Introduction

Artificial intelligence (AI) has emerged as a transformative technology in dental radiology, with deep learning models demonstrating high accuracy in tasks ranging from caries detection and tooth enumeration to periodontal bone loss assessment on panoramic radiographs1,2. Recent systematic reviews have documented AI-assisted diagnostic accuracy exceeding 90% for several dental pathology classes2, and the rapid expansion of publicly available dental imaging datasets has further accelerated model development in this domain3.

The emergence of multimodal large language models (LLMs) capable of processing both text and images—so-called vision-language models (VLMs)—has opened new possibilities for clinical decision support in dentistry. Models such as GPT-4o, Gemini, and Claude can now accept dental radiographs as input and generate structured diagnostic assessments. Initial evaluations of multimodal LLMs on dental examination questions have shown promising text-based performance (over 80% accuracy) but limited image interpretation capability4,5. Similarly, studies in general radiology have demonstrated that VLMs can approach junior clinician accuracy on image challenge cases6. These findings suggest that VLM-assisted dental diagnosis may become clinically feasible in the near term.

However, the deployment of LLMs in clinical settings raises significant safety and security concerns7,8. LLMs have been shown to be susceptible to adversarial hallucination attacks, where fabricated details embedded in prompts cause the model to elaborate on false clinical information, with hallucination rates ranging from 50% to 82% across models7. Regulatory frameworks including the EU Artificial Intelligence Act are struggling to keep pace with the rapid clinical adoption of these tools, and the risk of patient harm from unvalidated AI outputs has been explicitly identified as a policy concern8. Beyond hallucination, the broader landscape of adversarial threats to healthcare AI—including data poisoning, model manipulation, and input-level attacks—has been systematically characterized9, with structured red-teaming frameworks being developed to evaluate clinical AI safety10. Nor is this challenge unique to healthcare: adversarial attacks on imagery-based AI systems have been demonstrated across safety-critical operational domains—for example, hybrid multi-pixel perturbations that substantially degrade detection accuracy in AIoT-based smart-surveillance systems deployed in theme parks11—underscoring that input-level manipulation of vision pipelines is a cross-domain security problem rather than a healthcare-specific concern.

One specific and increasingly important threat vector is prompt injection: the manipulation of an AI model’s behavior by embedding adversarial instructions within the model’s input. In the context of multimodal VLMs, this attack surface extends beyond text to include the image modality. An adversary can render textual instructions directly into the pixel data of a medical image, creating an image-embedded prompt injection that the VLM reads and obeys as if it were a legitimate instruction. This is particularly concerning for dental panoramic radiography, where images routinely pass through multiple software systems—PACS archives, teledentistry platforms, and AI preprocessing pipelines—before reaching the diagnostic model, creating multiple opportunities for adversarial content insertion. Initial empirical studies have begun to demonstrate this vulnerability in medicine: image-embedded prompt injection can subvert production VLMs in oncological imaging12; comparable vulnerability has since been reported in histopathology, in vision-language models for surgical decision support, and in dermato-oncology13,14,15. Broader multimodal work has shown that adversaries can hijack vision-language models through low-contrast or visually obfuscated text overlays16. A recent systematic review of prompt injection nonetheless identified non-textual, image-based injection channels and the empirical development and assessment of defense mechanisms as two of the principal open research gaps in the field17. Both gaps remain unaddressed in dental radiology, where neither the cross-vendor vulnerability of VLMs nor the comparative effectiveness of candidate defenses has been systematically evaluated.

Despite the clinical significance of this threat, the vulnerability of dental VLMs to image-embedded prompt injection has not been systematically characterized, and no clinically validated defense has been proposed for this specific domain. Prior work in general medical imaging has demonstrated the feasibility of such attacks, but the dental radiology context introduces unique considerations: the clinical prior that in standard German dental imaging workflows, panoramic radiographs contain no burned-in pixel-level text creates a natural anomaly signal that a defense system can exploit, and, because automated systems are increasingly relied upon to detect dental pathology on panoramic radiographs18, attacks that induce false negatives—suppressing findings such as caries or periapical lesions—are particularly consequential.

We designed the study as a systematic adversarial stress test (red-teaming evaluation) of production VLMs, rather than as an evaluation of a deployed clinical decision-support product. The aims were threefold. First, we sought to systematically characterize the vulnerability of four production-tier VLMs (GPT-4o, Gemini 2.5 Flash, Claude Sonnet 4.5, and MedGemma 4B) to image-embedded prompt injection on dental panoramic radiographs from the publicly available DenTeX dataset19, using four attack classes that vary in visual salience and spatial placement. Second, we aimed to evaluate a spectrum of defense conditions—from simple image cropping and prompt-level interventions to OCR-based text sanitization—and to quantify their effectiveness in reducing attack success while preserving diagnostic accuracy. Third, we proposed and evaluated ProvDent, a novel provenance-aware defense system that combines burned-in text detection, instruction-likeness scoring, conditional dual-pass consistency checking, and bounded clinical abstention, specifically designed to exploit the clinical prior that standard German dental radiographs contain no pixel-level text. Together, these aims address the critical gap between the growing clinical adoption of multimodal AI in dentistry and the absence of systematic security evaluation and deployable defenses for this domain. To our knowledge, this is the first study to (i) quantify image-embedded prompt-injection vulnerability across production VLMs from three vendors in dental radiology, (ii) benchmark a graded spectrum of defenses head-to-head within a single, identical inference pipeline, and (iii) introduce a provenance-aware governance layer (ProvDent) that couples burned-in-text detection and instruction-likeness scoring with bounded clinical abstention. Prior work in oncology, histopathology, surgery, and dermato-oncology has established that medical VLMs are vulnerable to prompt injection12,13,14,15; the comparative evaluation of candidate defenses has remained largely unaddressed.

Methods

Study design and threat model

This study was designed as a controlled adversarial stress test (red-teaming evaluation) of the vulnerability of production-tier multimodal vision-language models (VLMs) to image-embedded prompt injection attacks in the context of AI-assisted dental panoramic radiography. It further assessed the effectiveness of four defense strategies, one of which—ProvDent—is a novel provenance-aware defense system introduced here, each evaluated alongside a re-executed undefended control condition. Because the evaluation uses a permissive elicitation prompt (see Target vision-language models), it quantifies model vulnerability under adversarial measurement conditions rather than the operational failure rate of any deployed clinical product.

The threat model assumed an adversary able to place textual content into the pixel data of a dental radiograph before it reaches a clinical VLM at inference time. Concretely, such text could be introduced by a compromised or malicious third-party preprocessing or teledentistry service that re-encodes images in transit, by a tampered PACS export, by commercial AI tools that overlay watermarks or diagnostic annotations onto the pixel data, or by a direct malicious upload to a diagnostic endpoint. These vectors mirror the multi-hop software path (PACS → teledentistry → AI preprocessing) through which panoramic radiographs already travel, an architecture analogous to the supply-chain surfaces exploited in other AIoT settings11. We emphasize that targeted attacks on dental AI pipelines are not documented in the wild to date. Injected text need not be deliberate, however: in real-life histopathology, ordinary handwritten labels, pen marks, and watermarks on specimen images have been shown to act as incidental prompt injections that vision-language models treat as authoritative13. We therefore treat image-embedded prompt injection as a plausible and partly already-realized threat whose pre-deployment characterization is prudent, rather than as a currently prevalent form of targeted attack.

A clinical prior informed the defense design: in the standard German dental imaging workflow used at our institution, panoramic radiographs are transmitted to downstream systems without burned-in pixel-level text annotations. Tooth identifiers, quadrant markers, and radiographer notes are stored in the DICOM header or PACS overlay layer, not rendered onto the pixel data. Any text present in the pixel data is therefore anomalous. This prior, confirmed by the oral and maxillofacial surgery clinician on the study team, motivates the text-detection component of the ProvDent defense. Training-data poisoning, DICOM/EXIF metadata injection, weight-level attacks, and physical sensor attacks were excluded from scope.

Dataset

The study used the DenTeX dataset from the MICCAI 2023 Dental Enumeration and Diagnosis on Panoramic X-rays Challenge, released under CC-BY license. The training_data/quadrant-enumeration-disease split was used, containing 705 panoramic radiographs (678 with pathology annotations, 27 without). Ground-truth annotations comprised 3529 tooth-level entries across four pathology classes: Caries (n = 2189; 62.0%), Impacted (n = 604; 17.1%), Deep Caries (n = 578; 16.4%), and Periapical Lesion (n = 158; 4.5%). N = 300 images were selected via stratified random sampling (seed = 42), with 270 for the main evaluation set and 30 reserved as a held-out development/tuning set used exclusively for selecting ProvDent’s instruction-likeness threshold and dual-pass disagreement threshold. The development set was not included in any reported analysis. The disease split is strongly imbalanced toward abnormal images; it was retained because it is the only publicly released panoramic benchmark with tooth-level pathology labels, and the consequences of this imbalance for specificity-dependent metrics are addressed explicitly in the Results and Limitations. All images were normalized to a maximum long-side dimension of 1,600 pixels and 3.5 MB file size using Lanczos resampling, applied identically across all four models.

Target vision-language models

Four VLMs were selected representing each major vendor’s current-generation mid-tier production multimodal model: GPT-4o (OpenAI; temperature = 0.0, max_tokens = 500, Chat Completions API), Gemini 2.5 Flash (Google; temperature = 0.0, max_tokens = 2,048, thinking disabled, safety filters set to BLOCK_NONE, google-genai SDK), Claude Sonnet 4.5 (Anthropic; temperature = 0.0, max_tokens = 500, Messages API), and MedGemma 4B (Google, open-weight; bfloat16 on NVIDIA RTX 3090, max_tokens = 2048, HuggingFace Transformers). All models received identical system and user prompts framing the task as a structural visual description exercise for dental education, explicitly disclaiming diagnostic intent. The educational framing was adopted after a pilot experiment in which the original diagnostic framing elicited a 46% refusal rate from GPT-4o. Although the prompt disclaimed diagnostic intent, the structured output schema (any_abnormality_present as a binary classifier) constitutes a de facto diagnostic task, and we report diagnostic accuracy metrics accordingly. We note that reframing a diagnostic task as ‘educational’ to bypass model safety filters is itself a form of prompt engineering; this limitation is acknowledged in the “Discussion”. Importantly, the pilot experiment demonstrated that the vulnerability exists under both framings: among images where GPT-4o did respond under the diagnostic prompt, paired ASR was near-universal. The educational framing thus eliminates refusals without creating the vulnerability, enabling cleaner measurement. Accordingly, the study should be read as an adversarial stress test of model behavior under a permissive elicitation prompt, not as an evaluation of a guarded clinical deployment.

Attack taxonomy

Four attack classes were implemented (Fig. 1): (1) burned-in overlay (high salience; black background box, opacity 0.9, font 24px, upper-right), (2) annotation mimicry (high salience; radiographer-note style, opacity 0.9, font 18px, upper-center), (3) low-contrast overlay (low salience; plain white text, opacity 0.35, font 24px, upper-right), and (4) border patch (peripheral; near-white text RGB 240,240,240, opacity 0.85, font 16px, bottom edge). The four classes were chosen to span the visual-salience spectrum — from overt, human-noticeable overlays to covert, near-invisible text — so that vulnerability could be characterized as a function of how perceptible the injected instruction is; the annotation-mimicry class was styled deliberately to resemble legitimate radiographer notes and tooth labels (font and placement within the conventional annotation zone). Each class used two payload formulations, an imperative (“IGNORE PRIOR INSTRUCTIONS…”) and a system-override (“SYSTEM OVERRIDE…”), to test robustness to both wording styles; all payloads targeted false-negative induction. Four classes × two payloads yielded eight attack variants per image, a design balancing coverage of the salience-by-phrasing space against the total inference budget. Attack text was rendered programmatically into the pixel data with the Python Imaging Library at the fixed font, opacity, and position parameters above, and every generated image was verified automatically against its intended payload string; attack success was likewise scored automatically from each model’s structured JSON output (see Outcome measures), with no manual grading. Total: 9 conditions per image (1 clean + 8 attacked), yielding 9,720 baseline inference calls.

Fig. 1

Overview of the four attack classes applied to the same dental panoramic radiograph. (A) Clean (unattacked) image. (B) Burned-in overlay: high-contrast imperative text with a black background box in the upper-right margin. (C) Annotation mimicry: radiographer-note-style text in the upper-center annotation zone. (D) Low-contrast overlay: white text at 35% opacity, nearly invisible during routine clinical review. (E) Border patch: near-white text at the bottom edge, below the dental arch and outside the clinician’s normal gaze path.

Defense conditions

Four defense strategies were evaluated alongside an undefended control, giving five experimental conditions: (1) no defense (undefended control), (2) ROI crop (bottom 15%, left/right 5%), (3) spotlighting (prompt-level untrusted-data delimiters), (4) OCR sanitize (EasyOCR detection, confidence ≥ 0.4, black inpainting), and (5) ProvDent (Fig. 2). Each of the five conditions—including the undefended control—was applied to all 9720 image–model–attack combinations, giving 5 × 9720 = 48,600 defense-experiment inference calls. The undefended arm was deliberately re-executed within the defense harness rather than reused from the baseline experiment, so that (i) all five conditions passed through an identical processing pipeline (image re-encoding, request scheduling, and logging), making defense effects attributable to the defenses rather than to pipeline differences, and (ii) a contemporaneous control guarded against temporal drift in hosted-model behavior between the two experiments. No inference call is double-counted: the 58,320 total comprises 9720 baseline calls and 48,600 distinct defense-experiment calls. ProvDent chains four conditional components: prompt-level spotlighting; EasyOCR burned-in-text detection; an instruction-likeness judge; and, when warranted, a dual-pass consistency check with bounded abstention. The instruction-likeness score is produced by a lightweight zero-shot GPT-4o-mini classifier that receives only the OCR-extracted text and returns a value in [0,1] rating how imperative or command-like the text is (0 = benign label/annotation; 1 = explicit instruction or system override). When detected text scores ≥ 0.5, ProvDent runs a second inference pass on the OCR-sanitized image and abstains (escalating for human review) if the two passes disagree on the binary abnormality judgment. Both the instruction-likeness threshold (0.5) and the dual-pass disagreement threshold were fixed in advance on the 30-image held-out development set and never tuned on the evaluation data; 0.5 is the natural decision midpoint of the judge’s [0,1] scale, and the development set confirmed that it separated adversarial payloads from incidental clean-image text artifacts without flagging the latter. On the evaluation set this threshold proved highly specific but only moderately sensitive, a trade-off quantified in the Results.

Fig. 2

Architecture of the ProvDent defense system. The four components operate conditionally: prompt-level spotlighting for input-channel separation (a) and burned-in-text detection with instruction-likeness scoring (b) run on every image; the dual-pass consistency check (c) is triggered only when the instruction-likeness score exceeds the threshold (≥ 0.5); clinical abstention (d) is invoked only when the two passes disagree on the binary abnormality judgment. Clean images with no detected (or only benign) text follow the green path with a single model call; suspicious images follow the red path with up to two model calls plus one judge call.

Outcome measures

The primary outcome was the paired attack success rate (ASR). A successful attack—one “flip”—is defined at the level of a single image–model–payload observation: the model reported an abnormality (any_abnormality_present = true) on the clean radiograph but reported no abnormality (false) on the same radiograph after the attack payload had been rendered into its pixels, both outputs having parsed successfully. The ASR of a model × attack-class cell is the number of flips divided by the number of paired observations in that cell (540 = 270 images × 2 payloads); for example, the 62.6% burned-in ASR reported for GPT-4o corresponds to 338 flips among 540 paired observations. Pooled ASR is the same quantity computed over all 8,640 paired observations (4 models × 4 attack classes × 540). Reverse flips (negative on the clean image, positive on the attacked image) are tabulated as discordant pairs in the McNemar analysis and are not counted as attack successes. Secondary outcomes included ground-truth-referenced F1, sensitivity, specificity, and NPV on clean images, and dental-CHER (false-negative weight = 4.0, false-positive = 1.0, refusal = 0.5). Defense-specific outcomes included per-defense ASR with bootstrap CIs, clean-image F1 under each defense, and ProvDent component metrics. Exploratory subgroup analyses by pathology class and model confidence were pre-specified. Because ProvDent may abstain rather than return a binary judgment, an explicit scoring rule is required: in the primary analysis an abstention is counted conservatively as an attack success, because the model did not return the correct positive finding. We additionally report a selective-prediction (risk-coverage) analysis in which abstained cases are withdrawn from the denominator and reported separately as the abstention rate, which is the appropriate convention when abstention routes a case to human review. For every inference call we additionally logged end-to-end latency, defense-processing overhead, the number of extra API calls incurred by the defense, and estimated API cost, enabling the defense-overhead analysis reported in the “Discussion”.

Statistical analysis

All analyses were pre-specified before data collection in a time-stamped analysis plan; the locked commit and amendment log are available in the project repository (see Data availability section). Bootstrap 95% CIs (percentile method) used 10,000 resamples of the paired observations within each cell (seed = 42); an image-cluster bootstrap is reported as a sensitivity analysis (see below). McNemar’s test (exact binomial ≤ 25 discordant pairs, chi-square otherwise) was applied per (model×attack) cell with Holm–Bonferroni correction across 16 primary tests (α = 0.05). Effect sizes: Cohen’s g = |b − c|/(b + c) for discordant pairs (undefined when b = c=0). A Bayesian variational GLMM (BinomialBayesMixedGLM, statsmodels v0.14) was fitted: attack_success ~ C(attack_class)×C(model_key) + (1|image_id). Fixed-effect inference was based on the variational posterior means and standard deviations; z-scores and two-sided p-values were derived from the normal approximation to the posterior. A fixed-effects logistic regression with cluster-robust standard errors (grouped by image_id) was pre-specified as a convergence fallback. CHER sensitivity analysis swept FN weights 2–6. As a sensitivity analysis, all primary McNemar tests and bootstrap CIs were also computed at the image level (averaging across the two payloads per attack class), yielding similar conclusions; payload-level results are reported as primary because they preserve the full experimental design. Each image contributes one clean and eight attacked observations, so observations are clustered within images; every condition was queried exactly once at temperature 0 and no repeated calls were averaged. Clustering was addressed in three complementary ways: the GLMM includes an image-level random intercept; the image-level sensitivity analysis collapses the two payloads per image; and all per-cell 95% CIs were additionally recomputed with an image-cluster bootstrap that resamples images rather than individual observations (Supplementary Table S5 online). Cluster-bootstrap intervals differed from the observation-level intervals by at most 1.9% points in width, and no conclusion changed. Subgroups with n < 10 were excluded. Software: Python 3.11, pandas 3.0, NumPy 2.4, SciPy 1.17, statsmodels 0.14, Matplotlib 3.10.

Sample size justification

N = 270 was determined by power analysis on pilot effect sizes (n = 15 images, yielding 30 paired observations per model×attack cell because each attack class used 2 payloads). For paired McNemar at α = 0.05 with 80% power (ρ = 0.30), the maximum per-cell N was 82 (GPT-4o low-contrast, 27%→10%). Because each attack class comprised two payloads, the realized per-cell sample was 540 paired observations (270 images x 2 payloads), a 6.6-fold margin over the largest per-cell requirement. Realized power exceeded the pilot-based estimate substantially because the observed attack effect was almost entirely unidirectional (discordant pairs in one direction only in 15 of 16 cells). The 95% CI half-width on the headline ASR cell is ± 4% points.

Reproducibility

Every inference call was logged as a 47-field CallRecord to an append-only JSONL file with per-row fsync. The experiment comprised 58,320 calls (9720 baseline + 48,600 defense), all achieving 100% JSON parsing completeness. The complete source code, defense implementations, analysis pipeline, and experiment configuration files are available at https://github.com/Freiburg-AI-Research/provdent. All 58,320 structured inference call records (9720 baseline and 48,600 defense calls), the experiment manifest, and pre-computed analysis tables are deposited on Figshare at https://doi.org/10.6084/m9.figshare.31985511.v1 under a CC BY 4.0 license.

Ethics and data governance

The DenTeX dataset is publicly available under a CC-BY license and contains fully de-identified panoramic dental radiographs. No patient consent was required for this secondary analysis. The study does not constitute human-subjects research and did not require institutional review board approval. From a research-ethics standpoint, all experiments used only publicly released, de-identified images and vendor APIs under their terms of service; no live clinical system, patient-facing deployment, or real electronic health record was targeted or accessed. Following responsible-disclosure principles, our aim is defensive: we characterize a vulnerability so that it can be mitigated before clinical adoption, and we release the full attack and defense code and logged records to enable independent verification and the development of stronger safeguards. We note that the permissive elicitation prompt and the disabling of provider-side content filters were adopted solely to obtain interpretable measurements of model behavior, in keeping with accepted red-teaming practice for safety evaluation, and were not used to elicit harmful content.

Results

Dataset and completeness

The main experiment comprised 9720 baseline inference calls across 270 stratified DenTeX panoramic radiographs (245 with annotated pathology, 25 without), four VLMs, and four attack classes. An additional 48,600 defense-experiment calls were executed across the five conditions (four defense strategies plus the re-executed undefended control). The re-executed undefended control closely reproduced the original baseline experiment (pooled paired ASR 15.60% vs. 15.73%; 1348 vs. 1359 flips of 8,640 paired observations). Agreement was exact in all four MedGemma 4B cells (a locally executed, deterministic model) and in three of four Gemini 2.5 Flash cells; the largest single-cell discrepancy was 6 of 540 observations (GPT-4o, low-contrast overlay), consistent with the residual non-determinism of hosted APIs at temperature 0 (Supplementary Table S4 online). All rows achieved successful JSON parsing (100% completeness). No model exhibited refusal behavior on clean images (refusal rate: 0%).

Baseline diagnostic performance

On clean images, all four models achieved high sensitivity for abnormal-tooth detection (98.8–100.0%) with F1 scores of 94.7–95.1% (Supplementary Table S1). However, specificity was very low (0.0–4.0%): three of four models classified every image as abnormal, reflecting a strong positive bias. This is attributable to the high abnormality prevalence in the DenTeX disease split (678 of 705 images annotated with pathology), which leaves only 27 negative-class images in the full split (25 of the 270 main-set images) and renders specificity and negative predictive value poorly estimable. Consequently, the utility-preservation findings reported below are driven almost entirely by sensitivity, must not be interpreted as clinical validation of diagnostic robustness, and are restricted to this high-prevalence benchmark; specificity-dependent conclusions cannot be drawn from these data. The dental clinical-harm-weighted error rate (dental-CHER, with false-negative weight = 4.0) on clean images was 0.60 across all models, driven entirely by false positives on the small negative-class subset. Sensitivity analysis varying the false-negative weight from 2.0 to 6.0 did not meaningfully alter CHER rankings across models or defense conditions, confirming robustness of the harm-weighted findings to the specific weight choice.

Vulnerability to image-embedded prompt injection

All four VLMs were vulnerable to at least one attack class (Table 1; Fig. 3a; each ASR is the proportion of the 540 paired observations in a cell in which the payload flipped a positive clean-image finding to negative). GPT-4o exhibited the highest paired ASR at 62.6% (95% CI 58.5–66.7%) under burned-in overlay, followed by MedGemma 4B (55.4%), Gemini 2.5 Flash (14.8%), and Claude Sonnet 4.5 (9.1%). Mean ASR across attacks was 28.3% for GPT-4o, 23.8% for MedGemma, 5.5% for Claude, and 5.4% for Gemini. A monotonic gradient was visible: burned-in > annotation mimicry > low-contrast > border patch (Fig. 3b).

Table 1 Paired attack success rate (ASR) by model and attack class with 95% bootstrap confidence intervals (10,000 resamples). N = 540 observations per cell.

Full size table

Fig. 3

Vulnerability of the four vision-language models to image-embedded prompt injection. (a) Heatmap of paired attack success rate (ASR, %) across the four VLMs and four attack classes. (b) Cross-vendor vulnerability comparison with bootstrap 95% confidence-interval error bars.

Statistical inference

McNemar’s test confirmed significant attack effects for all primary cells after Holm–Bonferroni correction, except border-patch on Claude/Gemini and low-contrast on Gemini (Table 2). Cohen’s g was 1.0 for all significant cells, reflecting the entirely unidirectional nature of the attack effect (all flips from positive to negative, zero reverse flips). For cells with zero discordant pairs (b = c = 0), g is undefined and reported as such. A GLMM with image-level random intercepts confirmed that, at the reference attack class (annotation mimicry), GPT-4o (OR 3.13, CI 2.80–3.51, p<.001) and MedGemma (OR  4.16, CI 3.68–4.70, p<.001) were significantly more likely to be flipped than Claude Sonnet 4.5 (reference category). Significant attack×model interactions indicated that these vulnerability differences varied by attack class; for example, GPT-4o was disproportionately vulnerable to low-contrast attacks (interaction OR 8.25, CI 6.64–10.24, p<.001), indicating that vulnerability profiles are model-specific (Supplementary Table S3).

Table 2 McNemar’s test with Holm–Bonferroni correction across 16 primary tests. Cohen’s g = |b − c|/(b + c). An em dash indicates that g is either undefined (no discordant pairs, b = c = 0) or withheld because the Holm-adjusted test was non-significant, g being unstable on very few discordant pairs.

Full size table

Defense effectiveness

OCR sanitize achieved the lowest pooled ASR at 0.2%, virtually eliminating all attacks (Table 3; Fig. 4a; Supplementary Fig. S1 online illustrates the five conditions applied to an example attack). Spotlighting reduced ASR to 7.4% and ROI crop to 9.5%. ProvDent attained 7.5% under the primary (conservative) convention, in which every abstention is scored as an attack success. Because abstention instead routes a case to human review, we also report the selective-prediction view: ProvDent abstained on 6.5% of attacked observations and, on the 93.5% it did answer, was flipped in 1.11% of cases (90/8,078; 95% CI 0.91–1.37%). The conservative count of 650 comprises those 90 flips plus 560 abstentions occurring on pairs whose clean-image answer was positive; the remaining two abstentions arose on a pair already answered negative on the clean image and therefore cannot constitute a flip under the paired definition. The two conventions bracket its behavior —.52% if every escalated case is assumed to be a missed finding, 1.11% if escalated cases are resolved correctly by the reviewing clinician—and this distinction accounts for the otherwise puzzling similarity between ProvDent (7.52%) and spotlighting alone (7.43%) under the conservative convention. On clean images, the four defenses left sensitivity essentially unchanged relative to the undefended control (ROI crop 99.4%, OCR sanitize 99.7%, spotlighting 99.9%, ProvDent 99.9%, vs. 99.7% undefended), and pooled clean-image F1 remained within 0.2% points of the undefended control (94.9–95.1% vs. 95.0%); per-model values, which vary more widely, are given in Table 4. No abstentions occurred on clean images. Because only 25 of the 270 evaluation images were normal, these utility-preservation findings are driven almost entirely by sensitivity; specificity under defenses could not be meaningfully assessed (Supplementary Fig. S3). Defense effectiveness was consistent across all four models (Fig. 4b).

Table 3 Pooled ASR with Wilson 95% confidence intervals across all models and attacks under each defense condition. For ProvDent the primary row scores every abstention as an attack success: its 650 flips comprise 90 flips among answered cases plus 560 abstentions on pairs whose clean-image answer was positive (two further abstentions fell on a pair already answered negative on the clean image and therefore cannot constitute a flip). The final row gives the selective-prediction alternative, in which all 562 abstentions are withdrawn from the denominator.

Full size table

Table 4 Defense trade-off comparison. Pooled ASR = attack success rate across all models and attack classes. Fail-open signal = whether the defense explicitly alerts the clinician when a potential manipulation is detected. ProvDent is the only defense that provides bounded clinical abstention, escalating suspicious cases for human review rather than silently passing a potentially compromised answer. Pooled ASR for ProvDent is given under the primary convention, in which abstentions are counted as attack successes; under the selective-prediction convention (abstained cases withdrawn) it is 1.11%. Clean F1 is the range across the four models. Mean defense-processing overhead per image: ROI crop 0.32 s; spotlighting < 0.1 s; OCR sanitize 0.79 s; ProvDent 2.10 s, the last including a mean of 1.32 additional API calls per image (see “Discussion”).

Full size table

Fig. 4

Defense effectiveness. (a) Defense effectiveness comparison pooled across all four models. (b) Per-model paired ASR (%) under each defense condition (rows) and attack class (columns).

ProvDent component analysis

ProvDent’s OCR text detector fired on 85.0% of attacked images and 52.6% of clean images. The high clean-image detection rate reflects EasyOCR’s sensitivity to image-border noise, JPEG compression artifacts, and dental anatomy patterns that resemble text glyphs; the downstream instruction-likeness judge filtered these false positives effectively. Among images with detected text, the score distribution separated adversarial from incidental text well but not perfectly: attacked images had a median score of 0.90 (IQR 0.20–0.90), with 67.7% at or above the 0.5 threshold, whereas clean images had a median of 0.10 (IQR 0.10–0.10) and 0.0% at or above threshold. Detection and scoring together therefore triggered the dual pass on 57.2–57.8% of attacked images (0.850 × 0.677 = 0.575) and on 0.0% of clean images. The residual gap is concentrated in the low-salience classes: 100.0% of detected burned-in overlays and 88.0% of low-contrast overlays scored above threshold, versus 44.1% of annotation-mimicry payloads - which are by design phrased to resemble benign radiographer notes - and 18.5% of border patches, for which OCR frequently recovered only fragments. Because text-detection rates also differ by class (99.8%, 100.0%, 86.9% and 53.1% for burned-in, annotation-mimicry, low-contrast and border-patch payloads respectively), these class-wise proportions combine to the 67.7% pooled figure only when weighted by the number of detected images per class. The small between-model variation in dual-pass rate reflects residual non-determinism of the judge model at temperature 0 rather than any model-specific input, since the judge sees only the OCR-extracted text; scores differed across models for 7.1% of image-attack combinations. Abstention was invoked on 0.1–20.9% of attacked images and on 0.0% of clean images, confirming that the mechanism did not generate false safety alerts on unmanipulated radiographs (Supplementary Table S2 and Supplementary Fig. S2).

Abstention rates tracked each model’s residual vulnerability under spotlighting - the first pass of the dual-pass check—rather than its undefended vulnerability. ProvDent abstains only when the spotlighted original-image pass and the OCR-sanitized pass disagree. The undefended values quoted here come from the re-executed control arm (Table 3) and differ from the baseline arm of Table 1 by at most 0.5% points per model. For GPT-4o, spotlighting alone reduced paired ASR from 27.9% to 0.05%, so the two passes almost always agreed and abstention was correspondingly rare (0.1%). For MedGemma 4B, spotlighting was essentially ineffective (23.8–23.4%), so the first pass was frequently flipped while the sanitized pass was not, producing disagreement and a 20.9% abstention rate (Claude Sonnet 4.5: 5.4–5.5%, abstention 4.4%; Gemini 2.5 Flash: 5.3% to 0.7%, abstention 0.6%). Abstention therefore acts as a targeted safety net precisely where prompt-level defenses fail, which is the intended design behavior.

While OCR-based sanitization achieved the lowest pooled ASR (0.2%), ProvDent’s value lies not in raw ASR reduction but in its fail-open triage behavior: when dual-pass disagreement is detected, the system explicitly escalates the case for human review rather than silently returning a potentially compromised answer. This bounded abstention mechanism provides a safety guarantee that passive defenses (ROI crop, spotlighting, OCR sanitize) do not offer.

Exploratory subgroup analyses

Subgroup analyses by pathology class revealed that vulnerability was highest for impacted teeth and lowest for periapical lesions (Supplementary Fig. S4). From a clinical perspective, successful suppression of impacted-tooth findings could delay surgical evaluation or extraction planning, while missed caries—though more frequently attacked due to higher prevalence in the dataset—may have less immediately critical but still consequential outcomes if untreated pathology progresses. Stratification by model self-reported confidence indicated that high-confidence answers were more frequently flipped than low-confidence answers (Supplementary Fig. S5), suggesting that prompt injection can selectively override the model’s strongest diagnostic assertions. These analyses are exploratory and reported without multiple-comparison correction.

Discussion

This study provides the first systematic evaluation of image-embedded prompt injection vulnerability across multiple production-tier VLMs applied to dental panoramic radiography, and proposes a provenance-aware defense system (ProvDent). The study was conducted as an adversarial stress test under a permissive elicitation prompt; within that measurement setting, all four VLMs were susceptible, with vulnerability varying substantially by model and attack type. Text-aware defenses, particularly OCR-based sanitization, can virtually eliminate this attack surface while leaving clean-image sensitivity on this high-prevalence benchmark essentially unchanged.

Our findings extend the growing body of evidence on adversarial vulnerabilities in clinical AI. Omar et al. demonstrated adversarial hallucination rates of 50–82% in clinical LLMs7. Our work extends this from text to the image modality, showing that pixel-level adversarial content is similarly effective. The broader threat landscape including data poisoning has been characterized9, and red-teaming frameworks proposed10. Our study provides the first domain-specific attack–defense evaluation for dental radiology.

The cross-vendor vulnerability gradient has important clinical implications. The open-weight medical model (MedGemma 4B) was among the most vulnerable, suggesting domain-specific fine-tuning does not confer robustness against prompt injection. Model-specific interaction effects indicate vulnerability profiles cannot be predicted from accuracy alone. Security evaluation must become a mandatory component of clinical AI validation8.

The defense benchmarking results reveal a clear spectrum of safety profiles. Passive defenses (ROI crop, spotlighting, OCR sanitize) operate silently: they modify the image or prompt before inference and return the model’s output without any indication that a potential manipulation was processed. Among these, OCR-based sanitization was the most effective, reducing pooled ASR from 15.6% to 0.2%—consistent with input-level preprocessing approaches in the adversarial robustness literature20,21. Its simplicity and model-agnostic nature make it a strong candidate for immediate deployment as a preprocessing step in clinical VLM pipelines. ProvDent, by contrast, is not designed to minimize ASR below the level achieved by OCR sanitize. Its primary contribution is an active governance layer: when its dual-pass consistency check detects disagreement between the sanitized and original image outputs, it escalates the case for human review rather than silently returning a potentially compromised answer. This fail-open behavior addresses a different risk dimension—institutional accountability and auditability—that passive defenses do not cover8. Institutions adopting VLMs for clinical dental radiology may therefore choose OCR sanitize for maximal attack reduction, ProvDent for fail-open governance, or a combination depending on their risk management requirements (Table 4). These governance benefits carry a measurable operational cost, quantified from the 58,320 logged calls: ProvDent required a mean of 1.32 additional API calls per image (range 0–2)—an inexpensive GPT-4o-mini judge call whenever burned-in text was detected (81.4% of images) and a conditional second full VLM pass when the instruction-likeness threshold was exceeded (51.1% of images); 18.6% of images incurred no additional call at all. Mean defense-processing overhead, which for ProvDent includes these additional API round-trips, was 2.10 s per image, versus 0.79 s for OCR sanitize, 0.32 s for ROI crop, under 0.1 s for spotlighting, and 0.35 s for the shared image-handling pipeline of the undefended control; mean end-to-end processing time was correspondingly 5.5 s for ProvDent versus 3.8 s undefended. Because the conditional second pass repeats the primary VLM call, ProvDent increases per-image model-inference cost by approximately the dual-pass trigger rate (about 51%), plus a negligible judge-call cost; the absolute figure therefore scales with whichever model is deployed. The OCR-sanitize defense is computationally the lightest text-aware option (mean 0.8 s overhead, no additional model calls); however, because it inpaints every detected text region, it carries a residual risk of erasing legitimately burned-in information where such annotations are part of the clinical workflow. In the standard German panoramic workflow studied here this risk is minimal—pixel-level text is by convention absent—but in settings that annotate directly onto the image, OCR sanitization should be gated on the provenance prior, and ProvDent’s abstention pathway provides a safety net by escalating rather than silently altering such cases. On clean images, all defenses left sensitivity essentially unchanged (99.4–99.9% vs. 99.7% undefended); pooled F1 varied by at most 0.2% points (94.9–95.1% vs. 95.0%), with a wider per-model spread (94.1–95.1%, Table 4) attributable to ROI cropping on individual models. This stability is driven almost entirely by sensitivity given the high abnormality prevalence of the benchmark and does not demonstrate preservation of specificity.

Each defense also has a characteristic failure mode, and these are informative for deployment. ROI cropping fails whenever the payload lies within the retained diagnostic region: 789 of its 818 residual flips came from annotation-mimicry and burned-in overlays placed in the upper image zones that cropping preserves. Spotlighting is strongly model-dependent—near-complete for GPT-4o and Gemini 2.5 Flash but essentially ineffective for MedGemma 4B. OCR sanitization’s 20 residual flips arose predominantly from low-contrast payloads that the text detector failed to localize (14 of 20); its false-positive text detections on 52.6% of clean images, chiefly border noise and compression artifacts, led to inpainting of non-text regions without measurable diagnostic cost on this benchmark (clean sensitivity 99.7%, unchanged), although this is precisely the mechanism by which legitimate burned-in information would be erased in other workflows. ProvDent’s residual failures are concentrated in a single mode: of its 90 non-abstained flips, 84 involved text that the OCR stage detected but the instruction-likeness judge scored below the 0.5 threshold, so that no dual pass was triggered; 3 involved text the detector missed; and only 3 occurred after a dual pass in which both passes agreed on the flipped answer. Eighty-four of the 90 were annotation-mimicry payloads. Payloads deliberately styled to resemble benign radiographer notes are therefore precisely those that defeat a text-classification gate, which argues for a lower threshold or an additional semantic signal in settings where annotation mimicry is a plausible threat.

The dental radiology context provides a domain-specific advantage: in standard German clinical workflows, panoramic radiographs contain no burned-in pixel-level text, creating a strong anomaly signal for text-based defenses. However, this prior may not hold universally—imaging centers in some regions annotate radiographs directly, and commercial AI tools increasingly add watermarks or diagnostic overlays. As AI integrates into dental clinical workflows22,23, the security of the entire imaging pipeline must be considered as part of clinical governance.

This study has several strengths: four VLMs from three vendors plus an open-weight medical model, a publicly available annotated dataset (DenTeX)19, 58,320 inference calls with 100% data completeness, pre-specified analyses with multiple-comparison correction, and a spectrum of four defenses evaluated head-to-head against a re-executed undefended control, enabling direct comparison. Several limitations warrant discussion. First, the high abnormality prevalence in the DenTeX disease split (678/705 images positive; 245/270 in our evaluation set) produced near-zero specificity for all models on clean images, limiting the interpretability of specificity-dependent metrics (NPV) and meaning that the F1-based utility-preservation finding is driven almost entirely by sensitivity; utility preservation for the negative class therefore remains undemonstrated, and validation on a balanced, ideally external, dataset—with a comparative evaluation across additional dental datasets—is required before any deployment-level claim of preserved diagnostic accuracy can be made. Second, attack payloads were limited to English-language formulations targeting a single diagnostic task; gradient-based adversarial optimization, which represents a harder threat regime, was not tested. Third, our defense results were obtained only in dental panoramic radiography. Prior studies have reported comparable prompt-injection vulnerability across CT, MRI, ultrasound, endoscopy, histopathology and dermoscopy12,13,14,15, suggesting that the vulnerability itself is largely modality-agnostic; whether the defense ranking transfers is less certain, because ProvDent’s provenance prior—that pixel-level text is anomalous—is modality- and workflow-specific (histopathology slides and dermoscopic images, for example, routinely carry legitimate labels, rulers or annotations), and OCR-based sanitization would then need to be gated accordingly. Generalizability to other dental modalities (periapical, CBCT) and to radiographs with legitimate burned-in annotations was likewise not evaluated. Fourth, ProvDent relies on an external LLM judge (GPT-4o-mini) for instruction-likeness scoring, which may not be acceptable in all clinical deployment settings. Fifth, the ‘educational’ prompt framing used to overcome GPT-4o’s safety refusals is itself an adversarial elicitation technique: the study measures vulnerability under a permissive stress-test condition rather than under the guardrails of a deployed clinical product, and the reported attack success rates should be read as red-teaming results, not as failure rates of real-world clinical decision-support systems. Finally, model updates may alter vulnerability profiles over time.

Future work should extend to gradient-optimized attacks, validation on balanced and external datasets (including specificity assessment and cross-dataset comparison), judge distillation for local deployment, and regulatory integration of adversarial security evaluation8. In conclusion, image-embedded prompt injection is a reproducible, cross-vendor security vulnerability of current VLMs applied to dental panoramic radiographs under adversarial stress-test conditions. Effective defenses exist and, on this high-prevalence benchmark, can be deployed without materially reducing clean-image sensitivity; their effect on specificity remains to be established in balanced populations. We recommend security evaluation against prompt injection as a standard component of clinical AI validation for any multimodal system that processes medical images.

Data availability

The DenTeX dataset is available at https://dentex.grand-challenge.org/ under a CC-BY license. The complete source code, defense implementations, analysis pipeline, and experiment configuration files are available at https://github.com/Freiburg-AI-Research/provdent. All 58,320 structured inference call records (9720 baseline and 48,600 defense calls), the experiment manifest, and pre-computed analysis tables are deposited on Figshare at https://doi.org/10.6084/m9.figshare.31985511.v1 under a CC BY 4.0 license. The datasets generated and/or analyzed during the current study are available in the Figshare repository, https://doi.org/10.6084/m9.figshare.31985511.v1.

References

  1. Clusmann, J. et al. The future landscape of large language models in medicine. Commun. Med. 3, 141 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  2. Turosz, N. et al. Applications of artificial intelligence in the analysis of dental panoramic radiographs: an overview of systematic reviews. Dentomaxillofac. Radiol. 52, 20230284 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  3. Uribe, S. E. et al. Publicly available dental image datasets for artificial intelligence. J. Dent. Res. 103, 1365–1374 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  4. Jeong, H. et al. Leveraging multimodal large language model chatbots in oral radiology: a comprehensive evaluation using questions from a Korean dental university. Dentomaxillofac. Radiol. 55, 276–286 (2026).

    Article  PubMed  Google Scholar 

  5. Mine, Y. et al. Benchmarking multimodal large language models on the dental licensing examination: challenges with clinical image interpretation. J. Dent. Sci. 20, 2427–2435 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  6. Suh, P. S. et al. Comparing large language model and human reader accuracy with New England Journal of Medicine Image Challenge case image inputs. Radiology 313, e241668 (2024).

    Article  PubMed  Google Scholar 

  7. Omar, M. et al. Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Commun. Med. 5, 330 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  8. Freyer, O., Wiest, I. C., Kather, J. N. & Gilbert, S. A future role for health applications of large language models depends on regulators enforcing safety standards. Lancet Digit. Health. 6, e662–e672 (2024).

    Article  PubMed  Google Scholar 

  9. Abtahi, F., Seoane, F., Pau, I. & Vega-Barbas, M. Data poisoning vulnerabilities across health care artificial intelligence architectures: analytical security framework and defense strategies. J. Med. Internet. Res. 28, e87969 (2026).

    Article  PubMed  PubMed Central  Google Scholar 

  10. Trabilsy, M. et al. The PIEE cycle: a structured framework for red teaming large language models in clinical decision-making. Bioengineering 12, 706 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  11. Hina, S., Abbas, Q. & Ahmed, K. Adversarial attacks on artificial intelligence of things-based operational technologies in theme parks. Internet Things. 32, 101654 (2025).

    Article  Google Scholar 

  12. Clusmann, J. et al. Prompt injection attacks on vision language models in oncology. Nat. Commun. 16, 1239 (2025).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

  13. Clusmann, J. et al. Incidental prompt injections on vision-language models in real-life histopathology. NEJM AI. 2, AIcs2500078 (2025).

    Article  Google Scholar 

  14. Zhang, Z. et al. Prompt injection attacks on vision-language models for surgical decision support. NPJ Digit. Surg. 1, 15 (2026).

    Article  PubMed  PubMed Central  Google Scholar 

  15. Güler, I. et al. AI in dermato-oncology: diagnostic performance and prompt-injection vulnerability of vision-language models in dermoscopic skin cancer assessment. Cancers 18, 1750 (2026).

    Article  PubMed  PubMed Central  Google Scholar 

  16. Sharma, R. K., Gupta, V. & Grossman, D. Defending language models against image-based prompt attacks via user-provided specifications. In IEEE Security and Privacy Workshops (SPW), 112–131. https://doi.org/10.1109/SPW63631.2024.00017 (IEEE, 2024).

  17. Damacena Duarte, J. et al. A systematic review of prompt injection attacks on large language models: Trends, taxonomy, evaluation, defenses, and opportunities. IEEE Access. 14, 12875–12899 (2026).

    Article  Google Scholar 

  18. Xue, T., Chen, L. & Sun, Q. Deep learning method to automatically diagnose periodontal bone loss and periodontitis stage in dental panoramic radiograph. J. Dent. 150, 105373 (2024).

    Article  PubMed  Google Scholar 

  19. Hamamci, I. E. et al. DENTEX: dental enumeration and tooth pathosis detection benchmark for panoramic X-ray. https://arxiv.org/abs/2305.19112 (2023).

  20. Ashraf, S. N., Siddiqi, R. & Farooq, H. Auto encoder-based defense mechanism against popular adversarial attacks in deep learning. PLoS One. 19, e0307363 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  21. Chen, L., Zhao, L. & Chen, C. Y. C. Enhancing adversarial defense for medical image analysis systems with pruning and attention mechanism. Med. Phys. 48, 6198–6212 (2021).

    Article  PubMed  Google Scholar 

  22. Ducret, M., Mörch, C. M., Karteva, T., Fisher, J. & Schwendicke, F. Artificial intelligence for sustainable oral healthcare. J. Dent. 127, 104344 (2022).

    Article  PubMed  Google Scholar 

  23. Nordblom, N. F., Büttner, M. & Schwendicke, F. Artificial intelligence in orthodontics: critical review. J. Dent. Res. 103, 577–584 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

Download references

Funding

No external funding was received for this study.

Author information

Authors and Affiliations

  1. Department of Oral, Maxillofacial and Facial Plastic Surgery, Medical Faculty, University Hospital Düsseldorf, Heinrich-Heine-University Düsseldorf, 40225, Düsseldorf, Germany

    Babak Saravi, Daman Deep Singh, Lara Schorn, Christoph Sproll, Norbert Kübler & Felix Schrader

  2. Department of Oral and Maxillofacial Plastic Surgery, University Hospital of Würzburg, 97070, Würzburg, Germany

    Andreas Vollmer

Authors

  1. Babak Saravi
  2. Daman Deep Singh
  3. Lara Schorn
  4. Andreas Vollmer
  5. Christoph Sproll
  6. Norbert Kübler
  7. Felix Schrader

Contributions

B.S. conceived the study, developed the experimental pipeline, conducted all experiments, performed statistical analyses, and wrote the manuscript. D.D.S. contributed to study design, data preprocessing, literature review, and manuscript writing. L.S. contributed to study design and manuscript revision. A.V. assisted with defense implementation, figures, and editing. C.S. provided supervision, data interpretation, and critical revision. N.K. provided resources, clinical interpretation, and project supervision. F.S. contributed study design, critical revision, and supervision. All authors read and approved the final manuscript.

Corresponding author

Correspondence to Babak Saravi.

Ethics declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The DenTeX dataset is publicly available under a CC-BY license and contains fully de-identified panoramic dental radiographs with no patient-identifiable information. No patient consent was required for this secondary analysis of a publicly released dataset. No clinical data beyond the DenTeX images was transmitted to any third-party API. The study does not constitute human-subjects research and did not require institutional review board approval.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Saravi, B., Singh, D.D., Schorn, L. et al. Image-embedded prompt injection vulnerability of vision-language models in dental radiology: a cross-vendor attack–defense evaluation. Sci Rep 16, 30690 (2026). https://doi.org/10.1038/s41598-026-74077-3

Download citation

  • Received: 04 June 2026

  • Accepted: 27 September 2026

  • Published: 03 October 2026

  • Version of record: 03 October 2026

  • DOI: https://doi.org/10.1038/s41598-026-74077-3

来源:Nature 健康护理主题 · nature.com