Keras versions prior to 3.15.0 contain a denial of service vulnerability when loading malicious .keras model files. The flaw arises from the lack of validation on dataset shapes and sizes during model loading, which can lead to unbounded memory allocation and out-of-memory crashes. This vulnerability allows attackers to cause the process to terminate by exhausting system memory. The issue affects machine learning pipelines that load untrusted models and bypasses previous fixes for related vulnerabilities. A patch is available in version 3.15.0.