Executive brief
ExecuTorch is Meta's runtime framework for executing machine learning models efficiently on edge devices. An integer overflow in the model loading process causes insufficient memory to be allocated, potentially allowing an attacker with a malicious model file to achieve code execution. This could compromise any application or device running ExecuTorch to load untrusted models.
Technical details
An integer overflow vulnerability exists in the memory allocation logic of ExecuTorch's model loader (runtime/core/memory_allocator.h). When loading a model file, insufficient bounds checking on size calculations causes a smaller-than-expected memory region to be allocated. The vulnerability is triggered when processing a crafted ExecuTorch model, and requires user interaction (loading a malicious model file). A successful exploit can result in heap buffer overflow, leading to code execution or information disclosure. The fix is available in commit 8f062d3f661e20bb19b24b767b9a9a46e8359f2b (adding overflow checks in the allocation logic).
Affected products
- Meta ExecuTorch prior to commit 8f062d3f661e20bb19b24b767b9a9a46e8359f2b (versions 0.1.0 through 0.6.0 affected)
Timeline
- 2025-08-08: disclosed
- 2025-08-08: patched: Fix available in commit 8f062d3f661e20bb19b24b767b9a9a46e8359f2b