llama.cpp CVE-2026-126974: a malicious GGUF file can trigger an integer overflow and run code; fixed in build b11972
- Update: run llama.cpp b11972 or later, and check bundled copies in apps and wrappers built on llama.cpp, which may lag upstream.
- Treat model files as code: load GGUF models, control vectors and multimodal projectors only from sources you trust, and verify hashes for files pulled from model hubs.
- Contain: run inference servers as unprivileged users or in containers without access to secrets.
Sources: Zero Day Initiative ZDI-26-975: llama.cpp GGUF File Parsing Integer Overflow Remote Code Execution Vulnerability (CVE-2026-126974, Nov 13, 2026) · llama.cpp PR #31336: common, mtmd: bound control vector layer indices and M-RoPE position loops (merged Oct 28, 2026) · llama.cpp release b11972 (first build containing the fix, Oct 28, 2026)