A landmark privacy audit published in Nature reveals that medical artificial intelligence systems are highly susceptible to near-perfect membership inference attacks, exposing severe privacy inequities across patient populations. As AI diagnostic tools become integral to clinical workflows for tasks ranging from chest X-ray interpretation to dermatological screening, their underlying training data remains vulnerable to exploitation. Researchers conducted the first patient-level privacy assessment across seven large-scale, real-world medical datasets, analyzing imaging and electronic health records for conditions including pneumonia, skin lesions, and cardiac anomalies. By training 200 independent models per patient and measuring prediction confidence variations, the team demonstrated that attackers can reliably determine whether specific individuals contributed to a model with minimal interaction. The study highlights a critical accuracy-privacy trade-off, indicating that as diagnostic models become more precise, their susceptibility to data extraction grows proportionally. Conventional privacy measures, such as pseudonymization or demographic averaging, proved ineffective against these targeted probes. Instead of treating privacy as a uniform metric, the researchers found that risk distribution is deeply skewed. Patients belonging to underrepresented demographics, including racial minorities, Medicaid recipients, and individuals with rare medical conditions, face significantly elevated exposure. Because the evaluation prioritized the highest risk score across all records submitted by a single patient, even a single vulnerable medical image effectively compromises individual identity. These findings underscore a systemic flaw in current AI security frameworks. Aggregate privacy benchmarks routinely obscure extreme individual risks, leaving marginalized cohorts disproportionately exposed. If these groups perceive medical AI as a threat to their confidentiality, willingness to contribute data will likely decline, ultimately degrading model performance and reinforcing existing healthcare disparities. To mitigate these vulnerabilities, the authors advocate for the implementation of patient-level differential privacy. This approach introduces mathematically rigorous noise into model training and outputs, ensuring that no individual record can be reverse-engineered, regardless of its rarity or sensitivity. The research establishes a new baseline for medical AI security, emphasizing that equitable privacy guarantees are essential to maintaining public trust and sustaining the iterative development of diagnostic algorithms.