CVE-2026-126976

llama.cpp audio projector metadata could overflow a buffer allocation

llama.cpp · llama.cpp

First documented disclosure · Review checkpoint

Summary

Opening a malicious model artifact could execute code in the model-loading process.

Severity and exploitation evidence

High · 7.0/10 · CVSS 3.0. Severity source.

Exploitation not established: Malicious exploitation remains unverified in the reviewed sources.

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

Audio multimodal projector dimensions could overflow an integer before buffer allocation.

Conditions required

A user opens an attacker-controlled GGUF artifact in an affected build.

Potential impact

Code can run with the current process’s permissions.

Affected versions and configurations

b8958 is confirmed affected; the full range is unspecified.

Fix and mitigation status

Correction e2e7a9b2d04e0404131e551243d4bd399ccfe606. A first packaged fixed build is unverified.

Practical next steps

Deploy a build verified to include the linked correction and load only trusted model artifacts.

Operating lesson

Treat model files as untrusted parser inputs requiring bounds and overflow checks.

Evidence and limits

The public correction PR opened June 18 at 22:49 UTC and explicitly describes audio allocation and integer-overflow checks; it merged June 19 as the commit linked by ZDI. The formal ZDI advisory followed November 13 and CVE publication November 17. ZDI supplies CVSS 3.0. The CNA confirms a tested build, not an exhaustive version interval.

Sources