CVE-2026-133710: llama.cpp b11003 before b12556 Heap Out-of-Bounds Write via hrm_text Model Loader
llama.cpp b11003 before b12556 contains an integer overflow in the hrmtext loader loadarchhparams() that allows attackers to bypass the block count check using crafted GGUF metadata. Attackers can supply a GGUF model whose layer factors wrap modulo 2^32, causing loadarchtensors() to write tensor pointers past model.layers, corrupting heap memory and potentially enabling code execution.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
llama.cppto a version that resolves this vulnerability.Fixed in b12556
Event History
Frequently Asked Questions
Which deployments are exposed?
Deployments using ggml-org llama.cpp versions from b11003 through b12555 are affected when they load hrm_text GGUF models. The issue is local in the supplied severity vector, but no privileges are required and user interaction is required.
What does an attacker need to exploit this?
An attacker needs to cause the target to load a crafted GGUF model containing malicious hrm_text metadata. The layer-factor values are chosen to wrap modulo 2^32, bypassing the loader's block-count check and causing writes beyond model.layers.
What should be done if patching is not immediately possible?
Do not load untrusted or externally supplied GGUF models that use the hrm_text architecture. Restrict model ingestion to validated, trusted files until an updated build is deployed.
How can I determine whether I may be affected?
Check whether your llama.cpp build is b11003 through b12555 and whether it loads hrm_text GGUF models. Review model sources and ingestion paths for untrusted GGUF files, especially files that can be supplied by users or obtained from external repositories.