CVE-2026-112107 Details

A heap buffer overflow vulnerability has been identified in ztxz16 fastllm versions through 1.8.2. The issue arises in the Safetensors Weight Loading component, specifically within the SafeTensorItem::CreateBuffer function in src/model.cpp. The vulnerability is caused by a mismatch between the declared shape and data_offsets of a tensor, allowing for an out-of-bounds write of 128 bytes into a heap allocation intended for only 4 bytes. This flaw can be exploited remotely by loading a malicious model directory that contains a crafted model.safetensors file. The vulnerability has been publicly disclosed and could potentially be exploited in the wild.

Users are advised to avoid loading model directories from untrusted sources with fastllm. The vulnerability can be fixed by adding a consistency check for the data_offsets and shape declarations in the Safetensors parser before any buffer allocation or data reading occurs.