GPU VulnDB

Database/AI/ML frameworks & serving

llama.cpp: integer overflow in hrm_text GGUF loader allows heap write past model layers

CVSS 8.5CVE-2026-133710AI/ML frameworks & servingcurated

Impact

A crafted GGUF model file sets layer factors that wrap modulo 2^32, which defeats the block count check in load_arch_hparams(). load_arch_tensors() then writes tensor pointers past the end of model.layers and corrupts heap memory. The record says this can potentially lead to code execution in the process that loads the model. On a shared inference fleet, if tenants or users can supply their own GGUF files to a llama.cpp-based server, a malicious model could run code in the serving process on the GPU node. The flaw is limited to the hrm_text architecture loader in builds b11003 through b12555.

Who can reach it

Anyone who can get a crafted GGUF model loaded by llama.cpp: a user who uploads or points the server at a model file, or an operator who pulls an untrusted model. No authentication to llama.cpp itself is needed, but someone has to load the file (CVSS AV:L, UI:P).

What to do

Upgrade llama.cpp to release b12556 or later, which adds the overflow check, and restart every server process that loads models. Tools that bundle their own copy of llama.cpp need a rebuild or a vendor update that includes the fix; the record does not list which downstream products are affected. Until you have patched, load GGUF models only from trusted sources.

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.