CVE-2026-115913: Python Pytorch Attention Probability Retained On Modulecve-2026-115913-python-pytorch-attention-probability-retained-on-module
An attention probability tensor (the softmax output over the O(T^2) score positions) computed during a forward pass is stored on the module itself (`self.$ATTR = ...`). nn.Module instances are long-lived and are shared across requests in an inference server, so this input-sized tensor is retained after the request finishes and sits outside per-request memory
