An unchecked integer overflow in ArmNN’s TensorShape::GetNumElements() causes the number of tensor elements to be under‑computed. During model optimisation, the resulting GetNumBytes() value is too small for the actual data size, and the BatchToSpaceNdLayer reads beyond the allocated heap buffer. This over‑read can expose data or crash the application, compromising confidentiality, integrity and availability. The core weakness is an unchecked integer overflow that propagates to a buffer over‑read, fitting CWE‑680, CWE‑126, and CWE‑190 patterns.

The flaw resides in the ArmNN Tensor implementation used for TensorFlow Lite parsing. Systems that load and optimise TFLite models with ArmNN, such as embedded ARM devices or server‑side inference pipelines, are affected. The vulnerability applies to all releases through the 2026‑03‑27 build of the library, as no specific fixed versions are identified.

The vulnerability has no public exploit score, and it is not listed in the CISA KEV catalog, but it has a high potential impact if a malicious TFLite file can be supplied to the host. The CVSS score is 6.2, indicating moderate severity. The EPSS score is <1%, suggesting low exploitation probability. Attackers who can control the model input can trigger the overflow, leading to a buffer over‑read. The attack likely requires local or network delivery of a crafted model file; no privileged escalation is believed necessary. Given the absence of existing patches, the risk remains significant for un‑updated installations.