qingyi26/GGUF_clip_memory_corruption

GGUF clip memory corruption PoC

This repository contains proof-of-concept GGUF files for a llama.cpp clip.cpp metadata type-confusion issue.

Included files:

• gguf-clip-short-both-u8.gguf
• gguf-clip-scalar-mean-u8.gguf
• gen_clip_oob_poc.py

Summary:

• clip.vision.image_mean and clip.vision.image_std are consumed as float[3]
• the loader does not verify array-ness, subtype, or minimum length before reading
• malformed GGUF metadata can therefore be interpreted with the wrong type during model load

Relevant code paths in llama.cpp:

• tools/mtmd/clip.cpp
• ggml/src/gguf.cpp

PoC notes:

• gguf-clip-short-both-u8.gguf stores image_mean and image_std as arr[u8,1]
• gguf-clip-scalar-mean-u8.gguf stores image_mean as scalar u8
• both files are valid enough to reach the vulnerable metadata-consumption path during load

This repo is intended for responsible vulnerability reporting and reproduction.