CVE-2026-133710

llama.cpp HRM-Text model metadata could overflow layer counts

ggml-org · llama.cpp · HRM-Text GGUF loader

First documented disclosure · Review checkpoint

Summary

Crafted model metadata could make tensor loading write beyond its layer array.

Severity and exploitation evidence

High · 8.5/10 · CVSS 4.0. Severity source.

Exploitation not established: Reviewed public sources describe the weakness, a correction or a controlled demonstration; malicious exploitation against users is not established.

Exploitation evidence source

Technical severity is not a measure of observed attacks or of risk to your deployment. Attempts, demonstrations, and successful compromise retain the qualifications in the source.

How the weakness works

A 32-bit product of layer factors wraps and passes the block-count check before pointer writes.

Conditions required

A user or service loads a malicious HRM-Text GGUF model.

Potential impact

Heap corruption and potential code execution in the loading process.

Affected versions and configurations

b11003 through b12555.

Fix and mitigation status

b12556, released November 19 at 23: 12: 54Z.

Practical next steps

Update to b12556 or later before loading untrusted HRM-Text models.

Operating lesson

Use checked dimensions and allocation bounds consistently throughout model loading.

Evidence and limits

Reviewed patch 50518b64 widens the product and checks layer bounds; release b12556 contains it. CNA datePublic is November 19, before its November 20 ingestion. CNA publication is 2026-11-20; this is distinct from the earliest reviewed public source and correction dates. VulnCheck assessments: CVSS 4.0 8.5; CVSS 3.1 7.8.

Sources