Integer Overflow Vulnerability in llama.cpp by ggml

CVE-2026-116829

What is CVE-2026-116829?

An integer overflow vulnerability exists in the gguf_init_from_reader function of llama.cpp versions prior to b11301. This flaw allows attackers to bypass validation checks for tensor element counts by supplying shapes that, due to their size, wrap to zero during processing. As a result, attackers can create GGUF files containing excessively large tensor extents that cause the ggml_nbytes() function to overflow, leading to out-of-bounds memory access when buffers are created using these compromised byte counts. This situation poses a significant risk to system integrity and security.

References