The vulnerability in ArmNN involves an integer overflow in the TensorShape::GetNumElements() function, which calculates the number of elements in a tensor by multiplying its dimensions using 32-bit unsigned arithmetic without checking for overflow. This can cause the function GetNumBytes() to return a smaller-than-actual allocation size. Consequently, during model optimization, specifically in the BatchToSpaceNdLayer within Optimize()->InferOutputShapes(), a heap-based buffer over-read can occur because the code reads beyond the allocated memory buffer. This vulnerability affects ArmNN versions up to 2026-03-27 and can be triggered by a specially crafted TensorFlow Lite (TFLite) model file. The CVSS 3.1 vector indicates the attack requires local access with low complexity and no privileges or user interaction, impacting availability but not confidentiality or integrity.

An attacker with local access can supply a malicious TFLite model file that triggers an integer overflow in tensor dimension calculations, causing a heap-based buffer over-read during model optimization. This can lead to application crashes or denial of service due to memory corruption. There is no indication of confidentiality or integrity impact. No known exploits are reported in the wild.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is available, avoid processing untrusted or malformed TFLite model files with vulnerable versions of ArmNN. Monitor vendor communications for updates and patches addressing this issue.