Executive brief
ExecuTorch is Meta's runtime library used for deploying AI models on edge devices and mobile platforms. An integer overflow vulnerability in the model loading process can cause memory corruption, allowing an attacker to execute arbitrary code on a device running a vulnerable version. This poses a significant risk to any application using ExecuTorch for machine learning inference.
Technical details
The vulnerability is an integer overflow (CWE-190) located in the HierarchicalAllocator::get_offset_address() function within ExecuTorch's memory management subsystem. When parsing specially crafted model files, integer arithmetic used to calculate memory offsets wraps around, causing objects to be written outside their allocated buffers. The attack vector is network-based with no authentication or user interaction required—an attacker can craft a malicious ExecuTorch model file and deliver it to a vulnerable target. Successful exploitation enables arbitrary code execution with full system privileges. The vulnerability was patched in commit 0830af8207240df8d7f35b984cdf8bc35d74fa73, released as version 0.7.0 across pip, Maven, and Swift package distributions.
Affected products
- Meta Platforms ExecuTorch < 0.7.0
- Meta Platforms ExecuTorch Android < 0.7.0
Timeline
- 2025-08-07: disclosed: Published by NVD
- 2025-08-08: advisory: Published to GitHub Advisory Database
- 2025-08-08: patched: Fixed in version 0.7.0 and commit 0830af8207240df8d7f35b984cdf8bc35d74fa73