The vulnerability exists in the __getitem__ method of the MultiModalDataset class within the docarray.data.torch_dataset module. The method uses user-supplied input (field) to dynamically access and modify attributes of an object. The field string is split by periods ('.') to create an access path. However, there is no validation or sanitization of this path. This allows an attacker to provide a path that traverses into the object's class hierarchy (e.g., using __class__, __base__, etc.) and modify attributes of fundamental Python objects. This is a classic prototype pollution vulnerability. The provided PoC demonstrates how an attacker can overwrite __subclasscheck__ on a metaclass, leading to a DoS when Pydantic models (commonly used with FastAPI, as shown in the PoC) are processed. The root cause is the lack of input sanitization on the field parameter before using it in getattr and setattr calls.