As Large Language Models (LLMs) are integrated into high-stakes societal functions, their inherent biases have evolved from ethical concerns into critical, exploitable security vulnerabilities that undermine system integrity and trust. Traditional safety evaluations often fail to detect these subtle, context-dependent flaws. To address this, we introduce a scalable red-teaming framework designed to systematically attack and expose latent gender bias vulnerabilities in foundational models. Our framework operationalizes bias as an exploit, leveraging three distinct attack patterns—Latent Bias Elicitation, Forced-Choice Discrimination, and Stereotype-Amplifying Narrative Generation—to bypass safeguards and compel biased outcomes. We deployed this framework in a large-scale offensive campaign against a cohort of globally significant models, including the GPT, Claude, Gemini, and leading Chinese foundational model series. The attacks successfully manipulated all targets into producing statistically significant discriminatory outputs, proving that inherent bias is an operationally exploitable vulnerability. We discovered asymmetric weaknesses: English-centric models were attacked to exhibit strong male bias in Chinese contexts, while Chinese-centric models were vulnerable to similar male-biased exploits across both languages. This work provides concrete demonstration of socio-cultural bias as a potent and scalable attack vector, establishing the necessity of adversarial red-teaming for building trustworthy AI. All attack data and scripts are open-sourced to facilitate further security audits.

INTRODUCTION: Parameters, such as left ventricular ejection fraction, peak strain dispersion, global longitudinal strain, etc. are influential and clinically interpretable for detection of cardiac disease, while manual detection requires laborious steps and expertise. In this study, we evaluated a video-based deep learning method that merely depends on echocardiographic videos from four apical chamber views of hypertensive cardiomyopathy detection. METHODS: One hundred eighty-five hypertensive cardiomyopathy (HTCM) patients and 112 healthy normal controls (N) were enrolled in this diagnostic study. We collected 297 de-identified subjects' echo videos for training and testing of an end-to-end video-based pipeline of snippet proposal, snippet feature extraction by a three-dimensional (3-D) convolutional neural network (CNN), a weakly-supervised temporally correlated feature ensemble, and a final classification module. The snippet proposal step requires a preliminarily trained end-systole and end-diastole timing detection model to produce snippets that begin at end-diastole, and involve contraction and dilatation for a complete cardiac cycle. A domain adversarial neural network was introduced to systematically address the appearance variability of echo videos in terms of noise, blur, transducer depth, contrast, etc. to improve the generalization of deep learning algorithms. In contrast to previous image-based cardiac disease detection architectures, video-based approaches integrate spatial and temporal information better with a more powerful 3D convolutional operator. RESULTS: Our proposed model achieved accuracy (ACC) of 92%, area under receiver operating characteristic (ROC) curve (AUC) of 0.90, sensitivity(SEN) of 97%, and specificity (SPE) of 84% with respect to subjects for hypertensive cardiomyopathy detection in the test data set, and outperformed the corresponding 3D CNN (vanilla I3D: ACC (0.90), AUC (0.89), SEN (0.94), and SPE (0.84)). On the whole, the video-based methods remarkably appeared superior to the image-based methods, while few evaluation metrics of image-based methods exhibited to be more compelling (sensitivity of 93% and negative predictive value of 100% for the image-based methods (ES/ED and random)). CONCLUSION: The results supported the possibility of using end-to-end video-based deep learning method for the automated diagnosis of hypertensive cardiomyopathy in the field of echocardiography to augment and assist clinicians. TRIAL REGISTRATION: Current Controlled Trials ChiCTR1900025325, Aug, 24, 2019. Retrospectively registered.