GPU VulnDB

Database/AI/ML frameworks & serving

llama.cpp: heap buffer overflow in GGUF file parsing lets a malicious model file run code

CVSS 7.8CVE-2026-126975AI/ML frameworks & servingcurated

Impact

llama.cpp's GGUF parser copies data into a heap buffer without checking its length first. A crafted model file can overflow that buffer, and ZDI says this can be used to run arbitrary code in the context of the loading process. On a GPU serving node, that means control of the inference process and whatever it can reach. Any setup that loads GGUF files supplied by users or pulled from outside sources is exposed. This flaw was fixed in the same pull request as CVE-2026-126974, but it is a different defect (a missing length check, not an integer overflow).

Who can reach it

Anyone who can get a crafted GGUF file loaded by llama.cpp, through a model upload, a model pull, or by getting an operator to open it. The scoring requires user interaction (loading the file) but no authentication.

What to do

Rebuild or update llama.cpp to a build that includes ggml-org/llama.cpp pull request 31336, then restart the processes that embed it. No reboot is needed. The record does not name a fixed release tag. Until then, load only trusted GGUF files.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.