llama.cpp b9975 tightens GGUF security by rejecting empty metadata keys
New release fixes a potential vulnerability in GGUF model files.
ggml-org rolled out llama.cpp b9975 on July 12, 2024, a minor but security-critical patch that rejects empty metadata keys in GGUF files. The GGUF format is the standard container for quantized LLMs used with llama.cpp. Empty metadata keys could allow malformed or malicious model files to slip past validation, potentially causing crashes or undefined behavior during inference. This fix closes that loophole by enforcing that every metadata key must be non-empty.
The release is available across all major platforms and backends: macOS (Apple Silicon with and without KleidiAI, Intel), iOS (XCFramework), Linux (x64/ARM64/s390x CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16), Windows (x64/ARM64 CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Android (ARM64 CPU), and openEuler (x86/ARM64 with ACL Graph). This broad support ensures that developers running local LLMs on anything from a Raspberry Pi to a multi-GPU workstation can update to a more robust parser.
- Rejects empty metadata keys in GGUF files (PR #24917) to prevent parsing errors or exploitation.
- Includes prebuilt binaries for macOS, Windows, Linux, Android, iOS, and openEuler across CPU, CUDA, Vulkan, ROCm, OpenVINO, SYCL, and HIP backends.
- Signed commit with GPG key B5690EEEBB952194 ensures authenticity of the release.
Why It Matters
Safer local LLM inference for millions of users running models via llama.cpp.