Integer Overflow Vulnerability in llama.cpp Model Loader by ggml-org
CVE-2026-133710
8.5HIGH
What is CVE-2026-133710?
The integer overflow vulnerability in the hrm_text loader of llama.cpp versions before b12556 allows attackers to exploit crafted GGUF metadata. By manipulating the layer factors to wrap modulo 2^32, attackers can bypass critical block count checks. This exploitation results in the load_arch_tensors() function potentially writing tensor pointers beyond allocated buffers, leading to heap memory corruption. Such corruption can enable an attacker to execute arbitrary code, posing a significant risk to systems utilizing affected versions of the llama.cpp library.
Affected Version(s)
llama.cpp b11003
